Current IssueSpecial IssueEditorial BoardOnline First ArticlesArchive
Current IssueSpecial IssueEditorial BoardOnline First ArticlesArchive

Full Research Article

Fecal-DANet: A Specialized Deep Learning Architecture for Automated Poultry Health Monitoring via Fecal Image Analysis

M
Malrey Lee1,*
0000-0002-1256-6795
1The Future Artificial Intelligence Technology Company, JeonJu 54893, Republic of Korea.

ABSTRACT

Background: Early detection of poultry diseases is critical for minimizing economic losses. While fecal analysis is standard, the lack of immediate specialist availability often delays diagnosis. Existing automated systems frequently struggle to distinguish between morphologically similar diseases.

Methods: We propose Fecal-DANet, which employs Decoupled Group Attention to precisely analyze structure and color. To handle varied image sizes, Multi-Kernel Group Convolutions were integrated. The model was trained on 8,420 high-resolution images categorized into Newcastle disease, Salmonellosis, Coccidiosis and Normal. Group Normalization ensured training stability.

Result: Fecal-DANet demonstrated superior performance in distinguishing easily confused conditions, such as Coccidiosis and Salmonellosis, by identifying subtle features like hemorrhagic mucus and bile discoloration. The model achieved a high F1-score and an overall accuracy of 98.89%, proving its reliability for early-stage farm diagnostics.

INTRODUCTION

Chicken is the most common poultry, yet diseases like coccidiosis cause severe nutritional deficiencies and global economic damage. As poultry demand for meat and eggs rises annually, automated disease diagnosis has become essential (Neethirajan, 2022; Shinde et al., 2026; Udhayavel et al., 2020). Traditional manual monitoring is increasingly limited by industrial scaling, necessitating innovative artificial intelligence (AI) solutions. Specifically, early detection of coccidiosis, salmonella and Newcastle disease is vital to prevent fatal productivity losses and massive economic impacts (Ajaykumar et al., 2024).
       
Diagnosis typically relies on changes in appearance, behavior and feces. This study focuses on fecal analysis, a widely adopted diagnostic proxy. Delays in diagnosis caused by a lack of on-site experts often lead to disease outbreaks, necessitating mass culling and antibiotic overuse, which incurs significant time and cost (Zelalem et al., 2023).
       
Various AI techniques have been explored to address these issues. While CNN-based models are common, they often struggle with local minima. Other approaches, including Faster R-CNN (Wang et al., 2019), VGG16 (Jyothi et al., 2025), Xception Ret (Mbelwa et al., 2021), Sequential CNN (Cho, 2025) and RBF-SVM (Okinda et al., 2019), have faced limitations such as limited datasets and poor discrimination of subtle pathological features. Furthermore, behavioral analysis methods are environmentally sensitive and often fail to capture critical diagnostic data.
       
To overcome these constraints, we propose Fecal-DANet, a specialized deep learning architecture for automated poultry health monitoring. The key contributions of this study are:

Unique architecture
 
Integration of Decoupled Group Attention (DGA) for independent chromatic and structural correction, combined with Multi-Kernel Group Convolution (MKGC) for multi-scale feature extraction.
 
Dataset refinement
 
Implementation of rigorous data cleaning to remove foreign substances, ensuring the model focuses exclusively on fecal disease indicators.
 
High performance
 
Achieving a 98.75% accuracy rate, particularly in distinguishing visually similar salmonellosis and coccidiosis through fine mucus and bloodstain distribution.
       
This system facilitates expert diagnosis and empowers non-professionals in remote areas, significantly enhancing the efficiency of precision livestock farming and early disease prevention.

MATERIALS AND METHODS

The experimental work was conducted at Furture Artificial Intelligence Tech, Jeonbuk National University during the period from January 2025 to December 2025.
 
Data set
 
The data used in this study was taken from Zenodo, an open source for studying poultry feces (Machuve et al., 2021).  The data consists of a total of 8,420 high resolution poultry feces images. Samples were classified into four classes: Coccidiosis, Newcastle Disease, Salmonella and Normal. In the model proposed in this study, all images went through a standardized preprocessing pipeline, including resizing and normalization, as shown in Fig  1, to enhance the feature extraction capability and allow the model to focus on key pathological indicators.

Fig 1: Standardized preprocessing pipeline of data.


       
Fig 1 shows the process of converting raw data into model input data through data cleaning, resizing and feature-specific normalization to ensure the stability and accuracy of the Fecal-DANet model. This pipeline is intended to enable the deep learning model to learn representative pathological characteristics of poultry rather than noise such as background noise or lighting changes in the data.
 
Original image acquisition
 
The fecal images used in this study were obtained from Zenodo, an open academic data source. This image data includes various backgrounds, noise and environmental elements of the location where the image was obtained. Additionally, when obtaining images with a camera, problems related to the environment, such as light intensity and illumination, are also included. These noises are disadvantageous factors in classifying diseases.
 
Data cleaning
 
The raw data taken at the farm is mixed with shadows due to changes in light brightness, blurred colors and surrounding environments such as feces on the floor, resulting in lower recognition rates. To solve this problem, a high resolution fecal image must be created through a purification process.
 
Consistent size adjustment
 
Fine traces of blood or mucus in the image of poultry will reduce the recognition rate. To solve this, we adjust the image to 224*224 pixels, which is the size with the best recognition rate.
 
Normalization
 
As the neural network layers of deep learning become deeper, the distribution of input data changes. Regularization is necessary to ensure that the proposed model training can reach a steady state.
       
The data refined through this process is used as high-resolution input data in the feature extraction or classification stage.
       
The data set in Table 1 is classified into normal, Newcastle disease, salmonellosis and cocidiasis and has a total of 8,420 pages. Classification was done in a balanced manner to avoid focusing on one disease. In other words, when evaluating the proposed model, it was designed so that the results did not favor only certain diseases. In addition, overfitting, where the model becomes too familiar with only the training data and reduces recognition performance for new data, was prevented. The data set ratio was 80% for training, 10% for validation and 10% for testing, which is ideal for deep learning. Through this distribution, Fecal-DANet can learn the color and morphological changes of feces for all infected diseases.

Table 1: Poultry fecal datasets.


       
Fig 2 shows the feces for each disease used in this study.

Fig 2: Samples of poultry fecal.


 
(1) Normal: The feces are hard, brown and mixed with clear white urate.
(2) Cocidosis: It has red, hemorrhagic mucus and a watery texture.
(3) Newcastle disease: It has a green bile color and contains a lot of moisture.
(4) Salmonellosis: Yellow mucoid excrement mixed with cream-colored excrement.
       
Since feces are not clearly distinguished, this study used a mixture of separate group attention (DGA) and multi-kernel group convolution (MKGC) to learn the characteristics of various colors and shapes in order to better recognize these fine details. Disjoint group attention (DGA) is a technique that independently and precisely extracts fine features of color or shape through an attention mechanism separated by group. Multi kernel group convolution (MKGC) is a technology that effectively recognizes lesions of various sizes (mucus, blood stains, etc.) at the same time by grouping and calculating kernels of various sizes.
 
Fecal-DANet
 
Fecal-DANet is engineered to enhance the recognition rate of minute pathological indicators in poultry feces. As illustrated in Fig 3, the architecture follows an encoder-decoder framework.

Fig 3: Architecture of Fecal-DANet.


 
Encoder (Feature extraction)
 
The encoder extracts complex fecal characteristics using Multi-Kernel Group Convolution (MKGC) blocks instead of standard layers. By processing 3 *3 and 5 * 5 kernels in parallel, the model simultaneously identifies obscure bleeding and diverse fecal discolorations. Group Normalization ensures training reliability by maintaining consistent neural network performance.
 
Decoupled group attention (DGA)
 
Positioned at the core of Fecal-DANet, the DGA module redefines encoded high-dimensional feature maps from a pathological perspective. DGA utilizes a decoupled structure that calculates channel and spatial attention through independent paths to suppress inter-channel interference. Channel attention amplifies disease-specific features, while spatial attention focuses on localized lesions. This mechanism enables the model to exclude environmental noise, such as contaminants or lighting, ensuring focused inference on fine fecal textures.
 
Decoder and skip connection
 
The decoder reconstructs compressed feature maps for optimal classification. Skip connections are applied to preserve high-resolution spatial information potentially lost during down-sampling. By bypassing detailed features directly to the decoder, the model distinguishes subtle texture differences without information loss.
 
Classification head
 
Data is summarized into feature vectors through Global Average Pooling (GAP). Finally, a Softmax activation function yields probability values for four categories: coccidiosis, Newcastle disease, salmonellosis and normal, outputting the highest probability as the final diagnosis.
 
Multi kernel group convolution (MKGC)
 
MKGC divides kernels into groups to capture features at various scales. Unlike standard convolutions, MKGC operates dual kernels (3 *3 and 5 *5) as shown in Fig 4(a). This allows the filter to identify both local details, such as minute blood spots and broader elements like overall fecal shape or color distribution. Group convolution enhances computational efficiency and prevents overfitting by limiting model complexity, distinguishing this approach from previous CNN-based studies.

Fig 4: Architectural of the proposed module: (a) MKGC (b) DGA.


 
Detachable group attention (DGA)
 
As shown in Fig 4(b), the DGA module separates spatial and channel information to refine feature maps. Given that poultry diseases present independent features-such as chromaticity (bile or blood) and morphology (mucus or wateriness)-DGA processes them through independent heads. The texture attention head recognizes fecal structures (hardness, wateriness), while the color attention head identifies specific pathological colors. This separation allows the model to diagnose diseases with high precision.
       
In the Fig 4(a), the generated feature maps are concatenated and then compressed using 1 * 1 convolutional layers to produce a unified representation.
       
Fig 4(b) is the structure of the DGA module. This module contains two independent attention branches that process texture related information and color related information separately. This separation strategy allows the model to highlight disease related visual indicators more effectively, ultimately improving the reliability of the classification results.
 
Experimental setup and training
 
In this section, we detail the computational environment and specific learning settings to ensure the reproducibility and technical validity of the results.
       
The proposed Fecal-DANet is implemented using the TensorFlow/Keras framework on the NVIDIA GPU environment, which provides the required parallel processing power for multi-kernel group convolution operations. One of the drawbacks of deep learning models is that they fall into local minima. Therefore, in order to optimize this Fecal-DANet, the learning rate is 0.0001 and a pathfinding algorithm that controls direction and speed was used. This is called Adam Optimizer.
       
In this study, Categorical Cross-entropy was adopted as the loss function to optimize multi-class classification performance. This led the model to learn the probability distribution for the four disease categories closer to the correct label.
       
In order to evaluate whether the learning model proposed in this study is a reliable model, an experimental evaluation was conducted by setting the training data at 80%, the verification data set at 10% and the test data set at 10% out of a total of 8,420 pieces of data. And if the learning does not work properly even after learning it 5 times, it is a waste of time and the performance does not improve, so we set it to stop learning at that time.
       
Table 2 shows the hyperparameters used when learning this model. The learning rate was optimized to 0.0001. These hyperparameters are generally known to be very effective against overfitting in deep learning, so they were also used in this model. This optimized learning rate ensures that the multi kernel group convolution does not exceed the global minimum of the loss function, enabling accurate learning and recognition of microscopic diseases such as mucus from salmonellosis in poultry and bleeding traces from cocidiosis.

Table 2: Hyperparameters for Fecal-DANet.


       
To enhance generalization ability and mitigate the risk of overfitting, we incorporated a dropout rate of 0.25 in fully connected layers (dense layers).
       
Moreover, in a deep learning model, if the image is too large, learning will be slow and if the image is too small, fine evidence such as traces of blood or mucus will be lost, so the balance of image resolution is very important. Therefore, this study used 224*224 resolution. As a result of experimental evaluation using these data, the final test result was 98.89%.

RESULTS AND DISCUSSION

Quantitative performance evaluation
 
The learning model of Fecal-DANet used 1,684 sheets as a test dataset. As a result, the premise test accuracy achieved a high performance of 98.89. Table 3 shows the precision, which is the accurate prediction ability, which represents the proportion of diseases predicted by this model to be coccidiosis, the recall, which is the ability of this learning model to find the disease without missing sample data with actual coccidiosis and the F1-score, which represents the data that maintains a balance between precision and recall.

Table 3: Classification performance of Fecal-DANet.


       
The model proposed in this study also used precision, recall and F1-score as indicators to verify performance, just as existing systems used for disease classification. And the final evaluation was made into Type 1, which is related to precision, indicating an error saying that the poultry is healthy but is diseased and Type 2, which is related to recall, which is judging the poultry to be diseased but healthy and just passing it by.
 
Precision
 
Indicates the proportion of actual diseases among those predicted by the learning model to be diseases. In normal classification of feces, a precision of 1 means that the learning model has never made an error. In other words, if the learning model diagnoses salmonella disease, this is the probability that it is true salmonella disease. In other words, it means careful study.
 
Recall rate
 
This is the rate at which the learning model finds all actual diseases without making a mistake. This is the actual probability of finding all diseased poultry without missing a single one. A recall rate of 0.99 means that all diseases have been found.
 
F1-score
 
This is a score that harmoniously combines precision and recall. This model obtained 0.98, which means that there is no error in diagnosing disease (precision) and almost all chickens with disease were found (recall rate). In other words, this means that it is a nearly perfectly balanced system.
       
This means that Fecal-DANet has excellent sensitivity in predicting that diseased feces are diseased and specificity in predicting that normal feces are normal.  The fact that the performance of the verification data and the learning data is very similar is clear evidence that the model did not simply memorize the entire problem it was trying to solve, but solved the problem by properly understanding the principles. In other words, this proves that the multi-kernel group convolution and separable group retention used in this model are very stable. Here, the 0.98 F1-score is evidence that this proposed learning model solves the local minima, a shortcoming of the existing CNN model.
       
Fig 5 shows that classification was completed through 40 iterations of learning. In other words, very little learning was done in the early stages of learning and the disease was perfectly learned and classified very quickly and stably during learning 40 times. A very narrow gap between the learning and validation curves means it is very good. In other words, the proposed model continuously changed the disease environment to prevent memorizing only specific diseases. As a result, we have the ability to predict diseases that are new to us. The experimental evaluation results showed a low loss value of 0.0684. This means that this number is almost identical to the correct answer. In other words, it is evidence of high reliability. Fig 6 shows the confusion matrix. This means that this model carefully analyzed which problem might be confused with which problem if the prediction was wrong. In other words, there may be a slight confusion between Salmonella and Newcastle disease, so this indicates a post-mortem analysis to supplement this part.

Fig 5: Training and validation accuracy/loss curve of Fecal-DANet.


       
As shown in Fig 6, Fecal-DANet has high sensitivity across all pathological classes. The overwhelming number of diagonal components shows that the model clearly distinguishes between the green bile staining of Newcastle disease or the yellow mucus texture of salmonellosis.

Fig 6: Confusion matrix of Fecal-DANet.


       
The results in Fig 6 are consistent with the values   shown in Table 3. In Fig 6, the diagonal line indicates an accurate diagnosis of the disease and other values   indicate errors. These low values mean that the model learned very well. Therefore, it is proof that the proposed learning model makes a very accurate diagnosis without making errors even in external environmental changes or symptoms of very similar diseases. Minor misclassifications also appeared in this learning model. For example, very mild confusion has arisen between cocidiosis and salmonellosis because the two diseases have clinical similarities due to the yellow mucoid stool and pigmentation commonly seen at certain stages of the infectious disease.
 
Analysis of model
 
The qualitative evaluation of the proposed model is explained. Fig 7 shows some of the cases diagnosed by Fecal-DANet by looking at feces. In the successful case shown in Fig 7, the green bile staining, a characteristic feature of Newcastle disease, was accurately identified and the bloody mucus of cocidiasis was also accurately diagnosed. These results indicate that the separation group attention (DGA) mechanism accurately recognized chromaticity and fecal texture characteristics. There have been cases of failure in this model as well, but it has been very rare. Number 5 is a case of diagnostic failure, the cause of which was failure to identify major disease signs when they were obscured by the environment in which poultry was raised, i.e. soil, litter, feathers, etc. This will be resolved in a way that can integrate data from various angles in future research.

Fig 7: Example of diagnose predictive.


 
Comparative analysis with existing studies
 
To benchmark the results of this study, the performance of Fecal-DANet was compared with previous AI-based poultry diagnostic models. As presented in Table 4, the proposed model outperformed traditional transfer learning and traditional CNN approaches.

Table 4: Performance comparison analysis.


       
Table 4 shows the overall accuracy rate of how the learning model diagnosed the disease.
       
Fecal-DANet outperformed existing CNN based approaches such as Faster R-CNN by 1 or the XceptionNet model proposed by 2. In particular, 3, 4,5 achieved a high accuracy of 94.32%, 97.80% and 98.83 %, but the system was limited by the need for complex hardware to track movement and posture features. On the other hand, this model achieved a better accuracy of 98.89% using only fecal images. We demonstrate that the integrated design of multi-kernel group convolution (MKGC) and separate attention proposed in this study can extract fine lesions and complex lesions that are easily missed by the standard architecture used by existing researchers. Therefore, if poultry farms use this diagnostic model, they can quickly and accurately make an accurate diagnosis, which can reduce significant losses and shorten the time for disease prevention.
       
Unlike conventional models that process color and shape simultaneously, Fecal-DANet separately analyzes chromaticity (e.g., bile green or blood red) and texture (e.g., wateriness and morphological traits). This specialized approach mimics expert diagnostic logic, resulting in a high overall accuracy of 98.89%.
       
The integration of Multi-Kernel Group Convolution (MKGC) allows the network to adapt to the inherent variability of fecal samples. Pathological markers vary significantly based on the poultry’s growth and infection stages. For instance, young chicks produce small feces with minute blood spots or mucus, requiring the precision of a 3*3 kernel. Conversely, mature poultry produces larger feces where symptoms like green discoloration or extensive mucus may cover the entire sample, necessitating a 5 *5 kernel for broader context. By utilizing multi-scale kernels, Fecal-DANet captures both subtle traces and global fecal patterns, surpassing standard CNN methods in learning accuracy.
       
Despite these results, real-world farm environments-where feces may be discolored by ammonia gas or obscured by bedding-present challenges. Future research will transition from static image diagnosis to temporal analysis, observing changes in feces over time to enhance diagnostic reliability in complex environments.

CONCLUSION

In this research, the fecal-DANet model was proposed and evaluated to increase the overall diagnostic accuracy rate in poultry disease diagnosis. The core of this model is not to use integrated synthesis, which is mainly used in existing machine learning techniques, but to integrate multi-kernel group convolution and separation group tension to recognize disease characteristics and even subtle saturation changes in feces of various sizes, thereby increasing accuracy. The results of the evaluation showed better results than previous studies. The overall accuracy rate is 98.89. In addition, it was confirmed that this model learning index focuses on key indicators of disease diagnosis such as bloody mucus and phlegm coloring.
       
Fecal-DANet’s model is very lightweight and efficient, making it well suited for the integration of real time monitoring systems within poultry farms.
       
This research can reduce the concerns of farmers and prevent the spread of diseases early by predicting diseases in wild animals in advance. Therefore, it can reduce the economic losses of the country and farms and help with national budget planning. You can also plan for prompt response from disease experts. And we can recommend medicine related to your disease. Accurately diagnosing a disease at an early stage can reduce unnecessary drug prescriptions. As a result, it can contribute to improving human health. In future research, we plan to expand a larger amount of data, conduct research with other parts of poultry as well as feces and diversify the types of poultry to create a standard model.

ACKNOWLEDGEMENT

None.
 
Disclaimers
 
The views and conclusions expressed in this article are solely those of the authors and do not necessarily represent the views of their affiliated institutions. The authors are responsible for the accuracy and completeness of the information provided, but do not accept any liability for any direct or indirect losses resulting from the use of this content.
 
Informed consent
 
All animal procedures for experiments were approved by the Committee of Experimental Animal Care and Handling Techniques of Jeonbuk National University.

CONFLICT OF INTEREST

The authors declare that there are no conflicts of interest regarding the publication of this article. No funding or sponsorship influenced the design of the study, data collection, analysis, decision to publish, or preparation of the manuscript.

REFERENCES


  1. Ajaykumar, K., Harishankar, K., Shri Rangasami, S.R., Saravanakumar, V., Yazhini, G., Rajanbabu, V. and Premalatha, K. (2024). Growth performance, quantitative analysis and economics of broiler chickens as influenced by herbal dietary additives as alternative growth booster. Indian Journal of Animal Research. 58(7): 1139-1147. doi: 10.18805/IJAR.B-5326.

  2. Cho, O.H. (2025). Exploration and implementation of computational convolutional neural network model for poultry disease detection. Indian Journal of Animal Research. 1-7. doi: 10.18805/IJAR.BF-1945.

  3. Jyothi, Ch.,  Bheemana, S.G.,  Srisai, P. and Chowhan, R. (2025). AI- powered Image-based poultry disease detection system. International Journal of Engineering Research and Science and Technology. 21: 3(1).

  4. Machuve, D., Nwankwo, E., Mduma, N., Mbelwa, H., Maguo, E. and Munisi, C. (2021). Machine Learning Dataset for Poultry Diseases Diagnostics. https://zenodo.org/records/462 8934. 

  5. Mbelwa, H., Mbelwa, J. and Machuve, D. (2021). Deep convolutional neural network for chicken diseases detection. International Journal of Advanced Computer Science and Applications. 12(2): 1-9. https://doi.org/10.14569/IJACSA.2021. 0120295.

  6. Okinda, C., Lu, M., Liu, L., Nyalala, I., Muneri, C., Wang, J., Zhang, H. and Shen, M. (2019). A machine vision system for early detection and prediction of sick birds: A broiler chicken model. Biosystems Engineering. 188: 229-242. https:// doi.org/10.1016/j.biosystemseng.2019.09.015.

  7. Udhayavel, S., Gopala Krishna Murthy, T.R., Gowthaman, V., and Senthilvel, K. (2020). Seasonal prevalence of poultry diseases in Namakkal District of Tamil Nadu, India. Indian Journal of Pure and Applied Biosciences. 8(2): 187- 194. doi: http://dx.doi.org/10.18782/2582-2845.8044.

  8. Shinde, R.S., Harshad, C. Chauhan, S., Patel, S.S., Sharma, K.K. et al. (2026). Molecular and culture-based diagnosis of Infectious Bursal Disease Virus (IBDV) in vaccinated and non-vaccinated poultry flocks. Indian Journal of Animal Research. 60(1): 130-134. doi: 10.18805/IJAR.B-4433.

  9. Neethirajan, S. (2022). Automated tracking systems for the assessment of farmed poultry. Animals 12(3): 232. https://doi.org/ 10.3390/ani12030232. 

  10. Wang, J., Shen, M., Liu, L., Xu, Y. and Okinda, C. (2019). Recognition and classification of broiler droppings based on deep convolutional neural network. Journal of Sensors. 1-10. https://doi.org/10.1155/2019/3823515.

  11. Zelalem, M. and Simegn, G. (2023). Smartphone based detection and classification of poultry diseases from chicken fecal images using deep learning techniques. Smart Agricultural Technology. 4: 100221. doi: 10.1016/j.atech.2023. 100221.
Published In
Indian Journal of Animal Research
Article Metrics
Metrics Icon
16
Views
0
Citations
Reviewed By
Share Article
Download Article
In this Article
    Contact us
    Follow us

    Full Research Article

    Fecal-DANet: A Specialized Deep Learning Architecture for Automated Poultry Health Monitoring via Fecal Image Analysis

    M
    Malrey Lee1,*
    0000-0002-1256-6795
    1The Future Artificial Intelligence Technology Company, JeonJu 54893, Republic of Korea.

    ABSTRACT

    Background: Early detection of poultry diseases is critical for minimizing economic losses. While fecal analysis is standard, the lack of immediate specialist availability often delays diagnosis. Existing automated systems frequently struggle to distinguish between morphologically similar diseases.

    Methods: We propose Fecal-DANet, which employs Decoupled Group Attention to precisely analyze structure and color. To handle varied image sizes, Multi-Kernel Group Convolutions were integrated. The model was trained on 8,420 high-resolution images categorized into Newcastle disease, Salmonellosis, Coccidiosis and Normal. Group Normalization ensured training stability.

    Result: Fecal-DANet demonstrated superior performance in distinguishing easily confused conditions, such as Coccidiosis and Salmonellosis, by identifying subtle features like hemorrhagic mucus and bile discoloration. The model achieved a high F1-score and an overall accuracy of 98.89%, proving its reliability for early-stage farm diagnostics.

    INTRODUCTION

    Chicken is the most common poultry, yet diseases like coccidiosis cause severe nutritional deficiencies and global economic damage. As poultry demand for meat and eggs rises annually, automated disease diagnosis has become essential (Neethirajan, 2022; Shinde et al., 2026; Udhayavel et al., 2020). Traditional manual monitoring is increasingly limited by industrial scaling, necessitating innovative artificial intelligence (AI) solutions. Specifically, early detection of coccidiosis, salmonella and Newcastle disease is vital to prevent fatal productivity losses and massive economic impacts (Ajaykumar et al., 2024).
           
    Diagnosis typically relies on changes in appearance, behavior and feces. This study focuses on fecal analysis, a widely adopted diagnostic proxy. Delays in diagnosis caused by a lack of on-site experts often lead to disease outbreaks, necessitating mass culling and antibiotic overuse, which incurs significant time and cost (Zelalem et al., 2023).
           
    Various AI techniques have been explored to address these issues. While CNN-based models are common, they often struggle with local minima. Other approaches, including Faster R-CNN (Wang et al., 2019), VGG16 (Jyothi et al., 2025), Xception Ret (Mbelwa et al., 2021), Sequential CNN (Cho, 2025) and RBF-SVM (Okinda et al., 2019), have faced limitations such as limited datasets and poor discrimination of subtle pathological features. Furthermore, behavioral analysis methods are environmentally sensitive and often fail to capture critical diagnostic data.
           
    To overcome these constraints, we propose Fecal-DANet, a specialized deep learning architecture for automated poultry health monitoring. The key contributions of this study are:

    Unique architecture
     
    Integration of Decoupled Group Attention (DGA) for independent chromatic and structural correction, combined with Multi-Kernel Group Convolution (MKGC) for multi-scale feature extraction.
     
    Dataset refinement
     
    Implementation of rigorous data cleaning to remove foreign substances, ensuring the model focuses exclusively on fecal disease indicators.
     
    High performance
     
    Achieving a 98.75% accuracy rate, particularly in distinguishing visually similar salmonellosis and coccidiosis through fine mucus and bloodstain distribution.
           
    This system facilitates expert diagnosis and empowers non-professionals in remote areas, significantly enhancing the efficiency of precision livestock farming and early disease prevention.

    MATERIALS AND METHODS

    The experimental work was conducted at Furture Artificial Intelligence Tech, Jeonbuk National University during the period from January 2025 to December 2025.
     
    Data set
     
    The data used in this study was taken from Zenodo, an open source for studying poultry feces (Machuve et al., 2021).  The data consists of a total of 8,420 high resolution poultry feces images. Samples were classified into four classes: Coccidiosis, Newcastle Disease, Salmonella and Normal. In the model proposed in this study, all images went through a standardized preprocessing pipeline, including resizing and normalization, as shown in Fig  1, to enhance the feature extraction capability and allow the model to focus on key pathological indicators.

    Fig 1: Standardized preprocessing pipeline of data.


           
    Fig 1 shows the process of converting raw data into model input data through data cleaning, resizing and feature-specific normalization to ensure the stability and accuracy of the Fecal-DANet model. This pipeline is intended to enable the deep learning model to learn representative pathological characteristics of poultry rather than noise such as background noise or lighting changes in the data.
     
    Original image acquisition
     
    The fecal images used in this study were obtained from Zenodo, an open academic data source. This image data includes various backgrounds, noise and environmental elements of the location where the image was obtained. Additionally, when obtaining images with a camera, problems related to the environment, such as light intensity and illumination, are also included. These noises are disadvantageous factors in classifying diseases.
     
    Data cleaning
     
    The raw data taken at the farm is mixed with shadows due to changes in light brightness, blurred colors and surrounding environments such as feces on the floor, resulting in lower recognition rates. To solve this problem, a high resolution fecal image must be created through a purification process.
     
    Consistent size adjustment
     
    Fine traces of blood or mucus in the image of poultry will reduce the recognition rate. To solve this, we adjust the image to 224*224 pixels, which is the size with the best recognition rate.
     
    Normalization
     
    As the neural network layers of deep learning become deeper, the distribution of input data changes. Regularization is necessary to ensure that the proposed model training can reach a steady state.
           
    The data refined through this process is used as high-resolution input data in the feature extraction or classification stage.
           
    The data set in Table 1 is classified into normal, Newcastle disease, salmonellosis and cocidiasis and has a total of 8,420 pages. Classification was done in a balanced manner to avoid focusing on one disease. In other words, when evaluating the proposed model, it was designed so that the results did not favor only certain diseases. In addition, overfitting, where the model becomes too familiar with only the training data and reduces recognition performance for new data, was prevented. The data set ratio was 80% for training, 10% for validation and 10% for testing, which is ideal for deep learning. Through this distribution, Fecal-DANet can learn the color and morphological changes of feces for all infected diseases.

    Table 1: Poultry fecal datasets.


           
    Fig 2 shows the feces for each disease used in this study.

    Fig 2: Samples of poultry fecal.


     
    (1) Normal: The feces are hard, brown and mixed with clear white urate.
    (2) Cocidosis: It has red, hemorrhagic mucus and a watery texture.
    (3) Newcastle disease: It has a green bile color and contains a lot of moisture.
    (4) Salmonellosis: Yellow mucoid excrement mixed with cream-colored excrement.
           
    Since feces are not clearly distinguished, this study used a mixture of separate group attention (DGA) and multi-kernel group convolution (MKGC) to learn the characteristics of various colors and shapes in order to better recognize these fine details. Disjoint group attention (DGA) is a technique that independently and precisely extracts fine features of color or shape through an attention mechanism separated by group. Multi kernel group convolution (MKGC) is a technology that effectively recognizes lesions of various sizes (mucus, blood stains, etc.) at the same time by grouping and calculating kernels of various sizes.
     
    Fecal-DANet
     
    Fecal-DANet is engineered to enhance the recognition rate of minute pathological indicators in poultry feces. As illustrated in Fig 3, the architecture follows an encoder-decoder framework.

    Fig 3: Architecture of Fecal-DANet.


     
    Encoder (Feature extraction)
     
    The encoder extracts complex fecal characteristics using Multi-Kernel Group Convolution (MKGC) blocks instead of standard layers. By processing 3 *3 and 5 * 5 kernels in parallel, the model simultaneously identifies obscure bleeding and diverse fecal discolorations. Group Normalization ensures training reliability by maintaining consistent neural network performance.
     
    Decoupled group attention (DGA)
     
    Positioned at the core of Fecal-DANet, the DGA module redefines encoded high-dimensional feature maps from a pathological perspective. DGA utilizes a decoupled structure that calculates channel and spatial attention through independent paths to suppress inter-channel interference. Channel attention amplifies disease-specific features, while spatial attention focuses on localized lesions. This mechanism enables the model to exclude environmental noise, such as contaminants or lighting, ensuring focused inference on fine fecal textures.
     
    Decoder and skip connection
     
    The decoder reconstructs compressed feature maps for optimal classification. Skip connections are applied to preserve high-resolution spatial information potentially lost during down-sampling. By bypassing detailed features directly to the decoder, the model distinguishes subtle texture differences without information loss.
     
    Classification head
     
    Data is summarized into feature vectors through Global Average Pooling (GAP). Finally, a Softmax activation function yields probability values for four categories: coccidiosis, Newcastle disease, salmonellosis and normal, outputting the highest probability as the final diagnosis.
     
    Multi kernel group convolution (MKGC)
     
    MKGC divides kernels into groups to capture features at various scales. Unlike standard convolutions, MKGC operates dual kernels (3 *3 and 5 *5) as shown in Fig 4(a). This allows the filter to identify both local details, such as minute blood spots and broader elements like overall fecal shape or color distribution. Group convolution enhances computational efficiency and prevents overfitting by limiting model complexity, distinguishing this approach from previous CNN-based studies.

    Fig 4: Architectural of the proposed module: (a) MKGC (b) DGA.


     
    Detachable group attention (DGA)
     
    As shown in Fig 4(b), the DGA module separates spatial and channel information to refine feature maps. Given that poultry diseases present independent features-such as chromaticity (bile or blood) and morphology (mucus or wateriness)-DGA processes them through independent heads. The texture attention head recognizes fecal structures (hardness, wateriness), while the color attention head identifies specific pathological colors. This separation allows the model to diagnose diseases with high precision.
           
    In the Fig 4(a), the generated feature maps are concatenated and then compressed using 1 * 1 convolutional layers to produce a unified representation.
           
    Fig 4(b) is the structure of the DGA module. This module contains two independent attention branches that process texture related information and color related information separately. This separation strategy allows the model to highlight disease related visual indicators more effectively, ultimately improving the reliability of the classification results.
     
    Experimental setup and training
     
    In this section, we detail the computational environment and specific learning settings to ensure the reproducibility and technical validity of the results.
           
    The proposed Fecal-DANet is implemented using the TensorFlow/Keras framework on the NVIDIA GPU environment, which provides the required parallel processing power for multi-kernel group convolution operations. One of the drawbacks of deep learning models is that they fall into local minima. Therefore, in order to optimize this Fecal-DANet, the learning rate is 0.0001 and a pathfinding algorithm that controls direction and speed was used. This is called Adam Optimizer.
           
    In this study, Categorical Cross-entropy was adopted as the loss function to optimize multi-class classification performance. This led the model to learn the probability distribution for the four disease categories closer to the correct label.
           
    In order to evaluate whether the learning model proposed in this study is a reliable model, an experimental evaluation was conducted by setting the training data at 80%, the verification data set at 10% and the test data set at 10% out of a total of 8,420 pieces of data. And if the learning does not work properly even after learning it 5 times, it is a waste of time and the performance does not improve, so we set it to stop learning at that time.
           
    Table 2 shows the hyperparameters used when learning this model. The learning rate was optimized to 0.0001. These hyperparameters are generally known to be very effective against overfitting in deep learning, so they were also used in this model. This optimized learning rate ensures that the multi kernel group convolution does not exceed the global minimum of the loss function, enabling accurate learning and recognition of microscopic diseases such as mucus from salmonellosis in poultry and bleeding traces from cocidiosis.

    Table 2: Hyperparameters for Fecal-DANet.


           
    To enhance generalization ability and mitigate the risk of overfitting, we incorporated a dropout rate of 0.25 in fully connected layers (dense layers).
           
    Moreover, in a deep learning model, if the image is too large, learning will be slow and if the image is too small, fine evidence such as traces of blood or mucus will be lost, so the balance of image resolution is very important. Therefore, this study used 224*224 resolution. As a result of experimental evaluation using these data, the final test result was 98.89%.

    RESULTS AND DISCUSSION

    Quantitative performance evaluation
     
    The learning model of Fecal-DANet used 1,684 sheets as a test dataset. As a result, the premise test accuracy achieved a high performance of 98.89. Table 3 shows the precision, which is the accurate prediction ability, which represents the proportion of diseases predicted by this model to be coccidiosis, the recall, which is the ability of this learning model to find the disease without missing sample data with actual coccidiosis and the F1-score, which represents the data that maintains a balance between precision and recall.

    Table 3: Classification performance of Fecal-DANet.


           
    The model proposed in this study also used precision, recall and F1-score as indicators to verify performance, just as existing systems used for disease classification. And the final evaluation was made into Type 1, which is related to precision, indicating an error saying that the poultry is healthy but is diseased and Type 2, which is related to recall, which is judging the poultry to be diseased but healthy and just passing it by.
     
    Precision
     
    Indicates the proportion of actual diseases among those predicted by the learning model to be diseases. In normal classification of feces, a precision of 1 means that the learning model has never made an error. In other words, if the learning model diagnoses salmonella disease, this is the probability that it is true salmonella disease. In other words, it means careful study.
     
    Recall rate
     
    This is the rate at which the learning model finds all actual diseases without making a mistake. This is the actual probability of finding all diseased poultry without missing a single one. A recall rate of 0.99 means that all diseases have been found.
     
    F1-score
     
    This is a score that harmoniously combines precision and recall. This model obtained 0.98, which means that there is no error in diagnosing disease (precision) and almost all chickens with disease were found (recall rate). In other words, this means that it is a nearly perfectly balanced system.
           
    This means that Fecal-DANet has excellent sensitivity in predicting that diseased feces are diseased and specificity in predicting that normal feces are normal.  The fact that the performance of the verification data and the learning data is very similar is clear evidence that the model did not simply memorize the entire problem it was trying to solve, but solved the problem by properly understanding the principles. In other words, this proves that the multi-kernel group convolution and separable group retention used in this model are very stable. Here, the 0.98 F1-score is evidence that this proposed learning model solves the local minima, a shortcoming of the existing CNN model.
           
    Fig 5 shows that classification was completed through 40 iterations of learning. In other words, very little learning was done in the early stages of learning and the disease was perfectly learned and classified very quickly and stably during learning 40 times. A very narrow gap between the learning and validation curves means it is very good. In other words, the proposed model continuously changed the disease environment to prevent memorizing only specific diseases. As a result, we have the ability to predict diseases that are new to us. The experimental evaluation results showed a low loss value of 0.0684. This means that this number is almost identical to the correct answer. In other words, it is evidence of high reliability. Fig 6 shows the confusion matrix. This means that this model carefully analyzed which problem might be confused with which problem if the prediction was wrong. In other words, there may be a slight confusion between Salmonella and Newcastle disease, so this indicates a post-mortem analysis to supplement this part.

    Fig 5: Training and validation accuracy/loss curve of Fecal-DANet.


           
    As shown in Fig 6, Fecal-DANet has high sensitivity across all pathological classes. The overwhelming number of diagonal components shows that the model clearly distinguishes between the green bile staining of Newcastle disease or the yellow mucus texture of salmonellosis.

    Fig 6: Confusion matrix of Fecal-DANet.


           
    The results in Fig 6 are consistent with the values   shown in Table 3. In Fig 6, the diagonal line indicates an accurate diagnosis of the disease and other values   indicate errors. These low values mean that the model learned very well. Therefore, it is proof that the proposed learning model makes a very accurate diagnosis without making errors even in external environmental changes or symptoms of very similar diseases. Minor misclassifications also appeared in this learning model. For example, very mild confusion has arisen between cocidiosis and salmonellosis because the two diseases have clinical similarities due to the yellow mucoid stool and pigmentation commonly seen at certain stages of the infectious disease.
     
    Analysis of model
     
    The qualitative evaluation of the proposed model is explained. Fig 7 shows some of the cases diagnosed by Fecal-DANet by looking at feces. In the successful case shown in Fig 7, the green bile staining, a characteristic feature of Newcastle disease, was accurately identified and the bloody mucus of cocidiasis was also accurately diagnosed. These results indicate that the separation group attention (DGA) mechanism accurately recognized chromaticity and fecal texture characteristics. There have been cases of failure in this model as well, but it has been very rare. Number 5 is a case of diagnostic failure, the cause of which was failure to identify major disease signs when they were obscured by the environment in which poultry was raised, i.e. soil, litter, feathers, etc. This will be resolved in a way that can integrate data from various angles in future research.

    Fig 7: Example of diagnose predictive.


     
    Comparative analysis with existing studies
     
    To benchmark the results of this study, the performance of Fecal-DANet was compared with previous AI-based poultry diagnostic models. As presented in Table 4, the proposed model outperformed traditional transfer learning and traditional CNN approaches.

    Table 4: Performance comparison analysis.


           
    Table 4 shows the overall accuracy rate of how the learning model diagnosed the disease.
           
    Fecal-DANet outperformed existing CNN based approaches such as Faster R-CNN by 1 or the XceptionNet model proposed by 2. In particular, 3, 4,5 achieved a high accuracy of 94.32%, 97.80% and 98.83 %, but the system was limited by the need for complex hardware to track movement and posture features. On the other hand, this model achieved a better accuracy of 98.89% using only fecal images. We demonstrate that the integrated design of multi-kernel group convolution (MKGC) and separate attention proposed in this study can extract fine lesions and complex lesions that are easily missed by the standard architecture used by existing researchers. Therefore, if poultry farms use this diagnostic model, they can quickly and accurately make an accurate diagnosis, which can reduce significant losses and shorten the time for disease prevention.
           
    Unlike conventional models that process color and shape simultaneously, Fecal-DANet separately analyzes chromaticity (e.g., bile green or blood red) and texture (e.g., wateriness and morphological traits). This specialized approach mimics expert diagnostic logic, resulting in a high overall accuracy of 98.89%.
           
    The integration of Multi-Kernel Group Convolution (MKGC) allows the network to adapt to the inherent variability of fecal samples. Pathological markers vary significantly based on the poultry’s growth and infection stages. For instance, young chicks produce small feces with minute blood spots or mucus, requiring the precision of a 3*3 kernel. Conversely, mature poultry produces larger feces where symptoms like green discoloration or extensive mucus may cover the entire sample, necessitating a 5 *5 kernel for broader context. By utilizing multi-scale kernels, Fecal-DANet captures both subtle traces and global fecal patterns, surpassing standard CNN methods in learning accuracy.
           
    Despite these results, real-world farm environments-where feces may be discolored by ammonia gas or obscured by bedding-present challenges. Future research will transition from static image diagnosis to temporal analysis, observing changes in feces over time to enhance diagnostic reliability in complex environments.

    CONCLUSION

    In this research, the fecal-DANet model was proposed and evaluated to increase the overall diagnostic accuracy rate in poultry disease diagnosis. The core of this model is not to use integrated synthesis, which is mainly used in existing machine learning techniques, but to integrate multi-kernel group convolution and separation group tension to recognize disease characteristics and even subtle saturation changes in feces of various sizes, thereby increasing accuracy. The results of the evaluation showed better results than previous studies. The overall accuracy rate is 98.89. In addition, it was confirmed that this model learning index focuses on key indicators of disease diagnosis such as bloody mucus and phlegm coloring.
           
    Fecal-DANet’s model is very lightweight and efficient, making it well suited for the integration of real time monitoring systems within poultry farms.
           
    This research can reduce the concerns of farmers and prevent the spread of diseases early by predicting diseases in wild animals in advance. Therefore, it can reduce the economic losses of the country and farms and help with national budget planning. You can also plan for prompt response from disease experts. And we can recommend medicine related to your disease. Accurately diagnosing a disease at an early stage can reduce unnecessary drug prescriptions. As a result, it can contribute to improving human health. In future research, we plan to expand a larger amount of data, conduct research with other parts of poultry as well as feces and diversify the types of poultry to create a standard model.

    ACKNOWLEDGEMENT

    None.
     
    Disclaimers
     
    The views and conclusions expressed in this article are solely those of the authors and do not necessarily represent the views of their affiliated institutions. The authors are responsible for the accuracy and completeness of the information provided, but do not accept any liability for any direct or indirect losses resulting from the use of this content.
     
    Informed consent
     
    All animal procedures for experiments were approved by the Committee of Experimental Animal Care and Handling Techniques of Jeonbuk National University.

    CONFLICT OF INTEREST

    The authors declare that there are no conflicts of interest regarding the publication of this article. No funding or sponsorship influenced the design of the study, data collection, analysis, decision to publish, or preparation of the manuscript.

    REFERENCES


    1. Ajaykumar, K., Harishankar, K., Shri Rangasami, S.R., Saravanakumar, V., Yazhini, G., Rajanbabu, V. and Premalatha, K. (2024). Growth performance, quantitative analysis and economics of broiler chickens as influenced by herbal dietary additives as alternative growth booster. Indian Journal of Animal Research. 58(7): 1139-1147. doi: 10.18805/IJAR.B-5326.

    2. Cho, O.H. (2025). Exploration and implementation of computational convolutional neural network model for poultry disease detection. Indian Journal of Animal Research. 1-7. doi: 10.18805/IJAR.BF-1945.

    3. Jyothi, Ch.,  Bheemana, S.G.,  Srisai, P. and Chowhan, R. (2025). AI- powered Image-based poultry disease detection system. International Journal of Engineering Research and Science and Technology. 21: 3(1).

    4. Machuve, D., Nwankwo, E., Mduma, N., Mbelwa, H., Maguo, E. and Munisi, C. (2021). Machine Learning Dataset for Poultry Diseases Diagnostics. https://zenodo.org/records/462 8934. 

    5. Mbelwa, H., Mbelwa, J. and Machuve, D. (2021). Deep convolutional neural network for chicken diseases detection. International Journal of Advanced Computer Science and Applications. 12(2): 1-9. https://doi.org/10.14569/IJACSA.2021. 0120295.

    6. Okinda, C., Lu, M., Liu, L., Nyalala, I., Muneri, C., Wang, J., Zhang, H. and Shen, M. (2019). A machine vision system for early detection and prediction of sick birds: A broiler chicken model. Biosystems Engineering. 188: 229-242. https:// doi.org/10.1016/j.biosystemseng.2019.09.015.

    7. Udhayavel, S., Gopala Krishna Murthy, T.R., Gowthaman, V., and Senthilvel, K. (2020). Seasonal prevalence of poultry diseases in Namakkal District of Tamil Nadu, India. Indian Journal of Pure and Applied Biosciences. 8(2): 187- 194. doi: http://dx.doi.org/10.18782/2582-2845.8044.

    8. Shinde, R.S., Harshad, C. Chauhan, S., Patel, S.S., Sharma, K.K. et al. (2026). Molecular and culture-based diagnosis of Infectious Bursal Disease Virus (IBDV) in vaccinated and non-vaccinated poultry flocks. Indian Journal of Animal Research. 60(1): 130-134. doi: 10.18805/IJAR.B-4433.

    9. Neethirajan, S. (2022). Automated tracking systems for the assessment of farmed poultry. Animals 12(3): 232. https://doi.org/ 10.3390/ani12030232. 

    10. Wang, J., Shen, M., Liu, L., Xu, Y. and Okinda, C. (2019). Recognition and classification of broiler droppings based on deep convolutional neural network. Journal of Sensors. 1-10. https://doi.org/10.1155/2019/3823515.

    11. Zelalem, M. and Simegn, G. (2023). Smartphone based detection and classification of poultry diseases from chicken fecal images using deep learning techniques. Smart Agricultural Technology. 4: 100221. doi: 10.1016/j.atech.2023. 100221.
    In this Article
      Contact us
      Follow us
      Published In
      Indian Journal of Animal Research

      Editorial Board

      View all (0)