DeepFowl: Disease prediction from chicken excreta images using deep learning

Shahina Anwarul; Tanupriya Choudhury; Ketan Kotecha

doi:10.1515/nleng-2024-0081

Article Open Access

DeepFowl: Disease prediction from chicken excreta images using deep learning

Shahina Anwarul , Tanupriya Choudhury and Ketan Kotecha

Published/Copyright: March 24, 2025

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Nonlinear Engineering Volume 14 Issue 1

Abstract

The poultry industry faces numerous obstacles in maintaining the health and well-being of the chicken population due to the widespread occurrence of various diseases. Early detection and timely intervention are crucial to prevent diseases and minimize losses. The advocate solution is the need of the hour to cater to the discussed challenges. Therefore, the authors proposed a novel method called “DeepFowl.” A modified DenseNet169 deep learning-based model has been utilized by DeepFowl to diagnose diseases by analyzing the images of chicken excreta. The DeepFowl method, enhanced by a proposed data oversampling algorithm, has improved disease recognition accuracy in chickens from 96.3 to 98.4%. The model’s performance was evaluated across metrics such as accuracy, recall, precision, and F1-score, consistently achieving a high value of 98.4%. This demonstrates its effectiveness in early disease detection among chickens. Additionally, DeepFowl supports several sustainable development goals (SDGs): SDG2 (food security and sustainable agriculture), SDG3 (public health), and SDG9 (innovation).

Keywords: DeepFowl; disease prediction; chicken excreta; poultry health; deep learning; sustainable development goals

1 Introduction

The global poultry sector is crucial for food security by providing protein and nutrition to millions [1]. However, it faces significant challenges from infectious diseases like avian flu and salmonellosis, which impact poultry health, productivity, and economic stability [2,3]. Disease outbreaks result in direct losses from mortality and reduced productivity, as well as indirect costs related to disease management and trade restrictions. Early detection of diseases is vital for effective intervention to minimize their spread [4]. Traditional methods of monitoring poultry diseases involve symptom observation, lab testing, and research but are often slow, resource-intensive, and require professional expertise [5]. Recent advancements in technology, particularly machine learning (ML) and computer vision offer promising solutions for enhancing disease monitoring [6,7,8]. Deep learning techniques such as convolutional neural networks (CNNs) can analyze visual data with high accuracy [9,10,11]. This study introduces DeepFowl, a novel method using an enhanced DenseNet169 deep-learning architecture to predict diseases from images of chicken droppings. Changes in fecal characteristics can indicate underlying health issues like infections [12]. By leveraging advanced technology to analyze these visual cues effectively, DeepFowl aims to address the challenges of maintaining poultry health more efficiently.

DeepFowl analyzes images of chicken excreta and extracts data about the health of poultry using deep learning techniques. There are steps to work on DeepFowl, such as collecting images from the available dataset, processing them, recognizing features using deep learning, and predicting diseases. CNN, trained on a labeled dataset of chicken excreta images to identify disease symptoms, is a key component of the DeepFowl technique. A model is developed in the proposed work by refining and optimizing the DenseNet deep learning model. Extensive experiments are conducted on a range of chicken excreta images obtained from Zenodo [13] to evaluate the efficacy of DeepFowl. The accuracy, recall, precision, and F1-score are the evaluation metrics used to assess the framework’s performance. The results indicate how strong and reliable DeepFowl is at predicting chicken diseases from excreta images, highlighting its usefulness for early disease detection and health monitoring in the poultry industry.

In summary, the objective of the proposed research is to provide an automated system that helps to provide a significant advancement in the field of data-driven early disease diagnosis and intervention in poultry health surveillance. Through deep learning and image analysis, DeepFowl can transform poultry health management practices, ultimately contributing to the sustainability and resilience of the poultry industry in the face of emerging health threats. This research endeavor achieves novelty by striking a balance between recognition accuracy and computational efficiency. In this research, a trade-off has been observed between accuracy and computation. Techniques that achieve high accuracy are expensive in terms of computational resources, while others prioritize computational efficiency at the cost of accuracy. The present research introduces a solution that overcomes this trade-off by presenting a computationally efficient model that maintains the state-of-the-art (SOTA) level of accuracy.

This research is organized into four sections. Section 1 provides a brief introduction and motivation for the proposed research. Section 2 discusses the materials and methods used in the study, including the proposed methodology, the dataset used, and the proposed augmentation algorithm. Section 3 illustrates the experiments conducted to assess the proposed approach. Finally, the conclusion and future scope of the intended research are discussed.

2 Materials and methods

The study introduces a framework for identifying chicken diseases. It utilizes chicken excreta images to predict the presence of three prevalent poultry diseases: Salmonella, New Castle Disease, and Coccidiosis. It also helps identify the healthy class. The classification of diseases is done by rectified DenseNet169.

2.1 Architecture of DenseNet

Huang et al. suggested an architecture for Dense CNNs (DenseNet) featuring convolutional and pooling layers, dense blocks (DB), transition layers (TL), and classification layers [14]. In this framework, convolution layers contain filters applied to the activation map, while pooling layers reduce the dimensionality of the activation map. DBs link all layers so that each layer receives input from all the previous layers, combining features and passing its activation maps to succeeding layers. This approach expands the channel count, resulting in a more intricate model. To manage this complexity, TLs are introduced. DenseNet also employs skip connections, similar to those in residual networks, to enhance network performance without increasing its depth [15]. These skip connections effectively mitigate the problem of vanishing gradients [16].

In a CNN, a composite function f ₁ helps to connect the adjacent layers. The function f ₁ includes convolution layers, batch normalization (BN), pooling layers, and ReLU. The previous layer’s output ¥ ȴ − 1 is used as the input to the successive layer ¥ ȴ , as shown in (1):

(1) ¥ ȴ = f 1 ( ¥ ȴ − 1 ) .

However, 0th, 1st, …, ( ȴ − 1 ) th layers in a DB are connected, such that the concatenated outputs [ ¥ 0 , ¥ 1 , ¥ 2 , … , ¥ ȴ − 1 ] from all previous layers are fed as input to the next layer. Eq. (2) represents the concatenated output of the DB. This design allows each layer to receive a collective input from all preceding layers, and hence, these networks are called dense convolutional networks, reflecting their dense connectivity:

(2) ¥ ȴ = f 1 ( [ ¥ 0 , ¥ 1 , ¥ 2 , … , ¥ ȴ − 1 ] ) .

2.2 Improved architecture of the DenseNet model

The present work introduced an improved version of DenseNet169 architecture by incorporating the concatenation of the results of the global max and average pooling [17], BN [18], and dropout [19] into the model’s classification layer. This research utilizes two pooling methods, max pooling and average pooling, described mathematically in (3) and (4).

(3) P = O max a , a ( ¥ ) ,

(4) P = O avg a , a ( ¥ ) .

Here, the max pooling operation is represented by O max a , a ( ¥ ) and the average (avg) pooling is represented by O avg a , a ( ¥ ) , both applied on the input activation map with dimensions of a × a . The input activation map from the preceding convolutional layer is represented by ¥ . The output P from the pooling layer is pivotal. It has been noted that optimal performance can sometimes be achieved by utilizing the max value of the activation map a H × a w × a C (where a H is the height, a w is the width, and a C is the count of channels in the activation map), while in other cases, the average value from the previous layer proves to be superior. To retain both values, we employed the “concatenate()” class available in Keras to merge the average and maximum values of the activation map.

Global pooling simplifies the classifier by converting each channel in a feature map into a single value, replacing densely connected or fully connected layers. BN is used to standardize the negative features coming from the previous convolutional layer, helping to deal with the challenge of covariate shift. Integrating dropout layers during model training helps mitigate overfitting. According to Garbin et al. [20], the model’s accuracy varies significantly based on the dropout rate. Hence, multiple values were experimentally evaluated to determine the most effective dropout value. In our approach, incorporating Leaky ReLU after the BN layer has proven beneficial, effectively addressing the issue of “dying ReLU.” Therefore, we adopted the Leaky ReLU activation function for improved results [21]. The value of Leaky ReLU can be calculated by:

(5) F ( t ) = max ( 0.01 × t , t ) .

The Leaky ReLU function returns t for positive inputs and a small value (specifically 0.01 times t) for negative inputs. This ensures non-zero outputs for negative values, maintaining a gradient on the left portion of the graph and effectively addressing the issue of dead neurons. Figure 1 illustrates the improved version of DenseNet169. The mathematical expression for this rectified architecture is given in (6)–(14):

(6) ¥ 6 = f 2 ( [ ¥ 0 , ¥ 1 , ¥ 2 , ¥ 3 , ¥ 4 , ¥ 5 ] ) ,

(7) ¥ 6 = TL ( ¥ 6 ) ,

(8) ¥ 18 = f 2 ( [ ¥ 6 , ¥ 7 , … , ¥ 17 ] ) ,

(9) ¥ 18 = TL ( ¥ 18 ) ,

(10) ¥ 50 = f 2 ( [ ¥ 18 , ¥ 19 , … , ¥ 49 ] ) ,

(11) ¥ 50 = TL ( ¥ 50 ) ,

(12) ¥ 82 = f 2 ( [ ¥ 50 , ¥ 51 , … , ¥ 81 ] ) ,

(13) O = Concat ( gap , gmp ) ,

(14) C = f 3 ( O ) .

Figure 1

The flow of the proposed methodology.

Here, [ ¥ 0 , ¥ 1 , … , ¥ 82 ] represents the combination of the outputs from layers within the DB. TL applied to the DB’s output includes BN, ReLU, a 1 × 1 convolutional layer, and global average pooling (gap). A composite function f 2 comprising BN, ReLU, a 1 × 1 convolution layer, followed by another BN, ReLU, and a 1 × 1 convolutional layer. The Concat operation signifies the merging of gap and global max pooling (gmp). The function f 3 incorporating BN, Leaky ReLU, dropout, and two densely connected (or fully connected) layers culminate in a logarithmic softmax function. The output of the function f 3 represents the classes’ count of the dataset (C).

2.3 Dataset used

The “ML Dataset for Poultry Diseases Diagnostics,” accessible from Zenodo.org, was utilized in this study [13]. This annotated dataset for diagnosing poultry diseases targets moderate-scale poultry farmers and includes images of chicken excreta. The images were gathered in Tanzania’s Arusha and Kilimanjaro regions over 5 months from 2020 September to 2021 February, using the Open Data Kit (ODK) mobile app. The dataset comprises images of normal fecal material categorized as the “healthy” class, along with fecal samples from chickens diagnosed with Coccidiosis under the “cocci” class.

Furthermore, fecal images were obtained 1 week post-Salmonella inoculation, representing the “salmo” class. Fecal images for the “ncd” class were collected from chickens inoculated with Newcastle disease within 3 days of inoculation. The dataset comprises a total of 8,067 images, categorized into four classes: cocci, healthy, ncd (Newcastle disease), and salmo. Among these, the cocci class contains 2,476 images, 2,404 images of the healthy class, 562 images of the ncd class, and 2,625 images of the salmo class. Figure 2 demonstrates the sample images from the dataset.

Figure 2

Sample images of the dataset.

2.4 Tweaking of hyperparameters

The act of tuning the hyperparameters of a machine-learning model is commonly known as hyperparameter tuning. Identifying the optimized hyperparameter values can enhance the accuracy, flexibility, and robustness of the model, resulting in improved performance on new data. The proposed model is optimized by selecting specific parameters and hyperparameters for thorough analysis and testing. Several hyperparameters, such as batch size, learning rate, image dimensions, number of epochs, and data oversampling, are used to fine-tune the proposed model. A high learning rate may cause overshooting of the value, while a low value may require a large number of iterations to reach the optimal value. Hence, we utilized the Learning Rate Finder Curve, a valuable tool for automatically determining an appropriate learning rate for a given model to establish the learning rate, as illustrated in Figure 3 [22]. The setting of hyperparameter values used for the experimental purpose is presented in Table 1.

Figure 3

Learning rate finder curve.

Table 1

Hyperparameter values

Hyperparameter	Value
Epoch count	10
Learning rate	10⁻³ (automatically calculated using learning rate finder curve)
Input size	120 × 120 × 3
Optimizer used	Adam
Batch size	32
Activation function	Leaky ReLU
Split ratio of dataset	80:20

2.5 Data augmentation

Data augmentation, also known as data oversampling, is an approach to increase the size of the dataset by creating versions of the images present in the dataset through image manipulation techniques. It helps improve model performance by diversifying and expanding the dataset [23]. If each class in the dataset contains s samples, with s varying across classes, the algorithm generates p samples per class to maintain a balanced dataset. The output image, I _out, is produced after applying transformations and is then added to the class C of the dataset. By implementing oversampling techniques, the dataset is balanced, ensuring an even distribution of images across all classes. The number of images used for oversampling is determined by identifying the class with the highest image count in the dataset. In the proposed study, Algorithm 1 is used to oversample the dataset. A single augmentation technique (t _i) containing r transformations from a given set T is applied to the randomly chosen image. Figure 4 represents the augmented images after applying the given algorithm.


Input: Dataset D comprises C distinct classes with each class containing at most p unique samples, where p represents the highest number of images in any class within the dataset. Additionally, T= [t1, t2, t3, t4, t5] denotes the set of transformations applied to the dataset.
t1 = [centerCrop + CLAHE + ShiftScaleRotate]
t2 = [ShiftScaleRotate + randomRotate90]
t3 = [resize + randomBrightness + flip]
t4 = [transpose]
t5 = [strongTransformation]
Output: Dataset D consists of C distinct classes each containing p unique samples.
function Oversampling (D, p, T)
	p ← target number of images per class
	for all C ϵ D do
		s ← number of images in class C
		while s ≤ p do
			I ← select one random image of C
			t _i ← select one random transformation function from T
			r ← transformations in t _i
			I ← I _out
			for all r ϵ t _i do
			I _out ← r(I _out)
			end for
			q ← I _out
			C ← C Ս q
			s ← s + 1
		end while
	end for
end function

Figure 4

Sample output of augmented images.

2.6 Evaluation metrics

In evaluating the outcome of the proposed DeepFowl framework and other SOTA models on the dataset under consideration, the authors employ three key evaluation metrics: accuracy, recall, precision, and the F1-score. Classification accuracy, denoted as Acc, is a fundamental metric calculated using (15). It assesses the overall accuracy of predictions made by the models. However, basing the selection of the best classifier solely on accuracy can be insufficient, especially in scenarios with imbalanced classes or unequal costs associated with false positives and false negatives. This constraint is referred to as the “accuracy paradox” [24]. To overcome the accuracy paradox challenge, we used precision, recall, and F1-score to get deeper insights into the model performance. In (16), precision assesses how well the model identifies instances among all instances predicted as positive. Recall, as described in (17), evaluates the model’s capability to correctly recognize all instances by comparing positives to all actual positive instances in the dataset. Precision and recall offer insights into model performance in tasks where accurately identifying positives is crucial. Furthermore, to combine precision and recall into a measure, the authors compute the F1-score, as outlined in (18). The F1-score is an assessment of the accuracy of the model considering both false positives and negatives through the harmonic mean of precision and recall:

(15) Accuracy ( Acc ) = Ṫ p + Ṫ n Ṫ p + Ḟ p + Ṫ n + Ḟ n ,

(16) Recall ( R ) = Ṫ p Ṫ p + Ḟ n ,

(17) Precision ( P ) = Ṫ n Ṫ n + Ḟ p ,

(18) F 1 -score ( F 1 ) = 2 × R × P R + P .

Here, Ṫ p is true positive and signifies that the model correctly predicts a positive class, Ṫ n is true negative, which means the model correctly identifies a negative class. Similarly, Ḟ p is a false positive, indicating that the classifier mistakenly judges the negative class as the positive class, and Ḟ n is a false negative, indicating that the classifier erroneously identifies the positive class as the negative class.

3 Experimental results and discussion

This research was conducted on the Windows 11 platform configured with a GPU environment featuring an NVIDIA GeForce GTX 1650 Ti with 4GB of memory. For the classification tasks focusing on poultry diseases, the Fastai library was utilized, leveraging its capabilities within the Python programming language. The experimental images were sourced from an open dataset available on Zenodo. The summary of the used dataset for the experiments is given in Table 2. This dataset comprises images categorized into four distinct classes relevant to poultry diseases. The experiments were conducted under varying conditions, specifically comparing results obtained with and without data augmentation techniques. The evolution of the intended model’s performance across epochs is analyzed in Tables 3 and 4. Additionally, Figure 5 presents the training and validation loss plotted against the processed batches, highlighting the model’s efficiency without signs of overfitting. Figure 6 illustrates the confusion matrix, offering a comprehensive view of classification performance. Furthermore, Figure 7 displays sample output results contrasting actual outcomes with model predictions, offering a concrete illustration of the model’s predictive accuracy. Together, these figures substantiate the robustness and effectiveness of our proposed approach. The model achieved an identification accuracy of 96.3% without utilizing data augmentation. However, by employing the proposed augmentation algorithm, the accuracy improved significantly to 98.4%. These findings unequivocally show how the augmentation algorithm enhances the model’s performance.

Table 2

Summary of dataset used for experiments

	Total images	No. of images of each class	Training	Validation
Augmentation	10,500	2,625	80% (i.e., 8,400 images)	20% (i.e., 2,100 images)
Without augmentation	8,067	Cocci: 2,476	80% (i.e., 6,454 images)	20% (i.e., 1,613 images)
		Healthy: 2,404
		NCD: 562
		Salmo: 2,625

Table 3

Performance of the model across each epoch considering augmentation

Epoch	Training loss	Validation loss	Accuracy	Precision	Recall	F1-score
0	0.637510	0.381271	0.864762	0.864858	0.864002	0.863725
1	0.401032	0.263219	0.909048	0.912421	0.908804	0.909390
2	0.299157	0.196252	0.933333	0.934098	0.933013	0.933300
3	0.249061	0.140624	0.951905	0.952776	0.951512	0.951786
4	0.218500	0.107745	0.964286	0.964504	0.964167	0.964175
5	0.162012	0.091352	0.968095	0.968270	0.967987	0.968096
6	0.132487	0.078208	0.976667	0.976689	0.976594	0.976701
7	0.143119	0.085537	0.971429	0.972188	0.971469	0.971564
8	0.107971	0.073903	0.975714	0.975887	0.975774	0.975824
9	0.091658	0.070739	0.981429	0.981510	0.981501	0.981504
10	0.094653	0.061885	0.984286	0.984407	0.984378	0.984387

Table 4

Performance of the model across each epoch without augmentation

Epoch	Training loss	Validation loss	Accuracy	Precision	Recall	F1-score
0	0.681473	0.403960	0.856789	0.822906	0.793906	0.805001
1	0.399114	0.273992	0.899566	0.873777	0.872590	0.873135
2	0.333123	0.235144	0.920645	0.914999	0.895456	0.903932
3	0.257664	0.177368	0.938624	0.940377	0.910419	0.923323
4	0.226968	0.176177	0.939244	0.942531	0.916611	0.928046
5	0.180048	0.167280	0.947293	0.947866	0.929330	0.937787
6	0.151301	0.148030	0.950403	0.953117	0.939584	0.945779
7	0.141682	0.142417	0.953503	0.948360	0.949168	0.948700
8	0.131420	0.120901	0.958462	0.957037	0.947939	0.952245
9	0.119775	0.113058	0.959702	0.962585	0.950294	0.956104
10	0.106217	0.107513	0.963422	0.962827	0.956129	0.959335

Figure 5

Variation in training and validation loss vs batches processed (a) with augmentation and (b) without augmentation.

Figure 6

Confusion matrix (a) with augmentation and (b) without augmentation.

Figure 7

Sample output images.

3.1 Comparative analysis of the proposed work with prior studies

The present study demonstrates remarkable efficacy, achieving an impressive accuracy of 98.4% within a mere 10 epochs, a notable achievement considering that existing literature typically requires a significantly larger number of epochs to attain similar results. Importantly, this performance was accomplished without the need for the target (i.e., no excreta localization using object detection is done) extraction, underscoring the efficiency and robustness of the proposed methodology. Table 5 provides a comprehensive comparison with other studies, highlighting the supremacy of our work over existing methods.

Table 5

Comparative analysis with existing work

Author and year	Dataset used	Methodology used	Accuracy (in %)	Precision (in %)	Recall (in %)	F1-score (in %)	No. of epochs
Yogi and Yadav [25]	Zenodo open dataset	EfficientNetB7	97.07	99.7	99.7	99.7	120
Degu and Simegn [26]	Zenodo open dataset	YOLO for object localization + ResNet50	98.7	98.72	98.71	98.71	100
Wang et al. [27]	Self-curated dataset of 10,000 images	Faster R-CNN	NR	93.3	99.1	NR	50,000
Aziz and Othman [28]	Self-curated dataset	GLCM + SVM	93.75	NR	NR	NR	NR
Proposed work	Zenodo open dataset	Rectified DenseNet169	98.4	98.4	98.4	98.4	10

*NR: Not reported.

4 Conclusion and future scope

The conducted experiments on poultry disease diagnostics using fecal images have demonstrated the efficacy of the proposed work in identifying various poultry diseases. Utilizing a dataset of fecal images annotated with different disease classes, the experiments were performed on NVIDIA GeForce GTX 1650 Ti with 4GB of memory, leveraging the Fastai library for classification tasks. The results indicate that the proposed model can accurately discriminate healthy and diseased samples, specifically identifying conditions such as Coccidiosis, Salmonella, and Newcastle disease. This method presents a tool for moderate-scale poultry farmers offering a cost-effective and efficient approach for early detection and control of chicken diseases.

However, there are opportunities for improvements and further research. Future efforts could focus on expanding the dataset to cover a range of diseases and environmental conditions, which could enhance the model’s robustness and practicality. Another area worth exploring is implementing real-time disease detection systems by integrating trained models into mobile applications that can provide diagnostics in the field. Additionally, investigating deep learning architectures and comparing their performance with the model could provide valuable insights into optimizing disease detection accuracy. Collaborating with farmers and veterinarians to continuously validate and update the model with data could also be advantageous. By addressing these research directions, the effectiveness and utility of deep learning-based disease diagnosis in poultry farming can be significantly improved, resulting in better disease control practices and healthier poultry populations.

sanwarul@ddn.upes.ac.in

Acknowledgments

We are grateful to the Editor and anonymous reviewers for their valuable feedback. We thank both the Universities: UPES, Dehradun, Uttarakhand, India and Symbiosis Institute of Technology, Symbiosis International (Deemed University) (SIU), Pune, Maharashtra, India for jointly collaborating for this particular research work.

Funding information: We acknowledge the funding received from RSF (Research Support Funding) from Symbiosis Institute of Technology, Symbiosis International (Deemed University) (SIU), Lavale Campus, Pune, Maharashtra, 412115, India.
Author contributions: Shahina Anwarul, Tanupriya Choudhury, and Ketan Kotecha performed the implementation and wrote the introduction, literature review, methodology, and results sections of the manuscript. Finally, all authors reviewed the complete manuscript. All authors have accepted responsibility for the entire content of this manuscript and approved its submission.
Conflict of interest: Authors state no conflict of interest.
Data availability statement: All data generated or analyzed during this study are included in this published article.

References

[1] Birhanu MY, Osei-Amponsah R, Yeboah Obese F, Dessie T. Smallholder poultry production in the context of increasing global food prices: roles in poverty reduction and food security. Anim Front. 2023 Feb;13(1):17–25.10.1093/af/vfac069Search in Google Scholar PubMed PubMed Central

[2] Swayne DE, Akey BL. Avian influenza control strategies in the United States of America. In Avian Influenza 2005. Dordrecht: Springer; p. 113–30.10.1007/1-4020-3441-5_13Search in Google Scholar

[3] Jones PJ, Niemi J, Christensen JP, Tranter RB, Bennett RM. A review of the financial impact of production diseases in poultry production systems. Anim Prod Sci. 2018 Oct;59(9):1585–97.10.1071/AN18281Search in Google Scholar

[4] Astill J, Dara RA, Fraser ED, Sharif S. Detecting and predicting emerging disease in poultry with the implementation of new technologies and big data: A focus on avian influenza virus. Front Vet Sci. 2018 Oct;5:263.10.3389/fvets.2018.00263Search in Google Scholar PubMed PubMed Central

[5] Mremi IR, Rumisha SF, Sindato C, Kimera SI, Mboera LE. Comparative assessment of the human and animal health surveillance systems in Tanzania: Opportunities for an integrated one health surveillance platform. Glob Public Health. 2023 Jan;18(1):2110921.10.1080/17441692.2022.2110921Search in Google Scholar PubMed

[6] Subramani T, Jeganathan V, Kunkuma Balasubramanian S. Machine learning and deep learning techniques for poultry tasks management: a review. Multimed Tools Appl. 2024 Mar;84:1–43.10.1007/s11042-024-18951-0Search in Google Scholar

[7] Zhang L, Guo W, Lv C, Guo M, Yang M, Fu Q, et al. Advancements in artificial intelligence technology for improving animal welfare: Current applications and research progress. Anim Res One Health. 2024 Feb;2(1):93–109.10.1002/aro2.44Search in Google Scholar

[8] Wurtz K, Camerlink I, D’Eath RB, Fernández AP, Norton T, Steibel J, et al. Recording behaviour of indoor-housed farm animals automatically using machine vision technology: A systematic review. PLoS one. 2019 Dec;14(12):e0226669.10.1371/journal.pone.0226669Search in Google Scholar PubMed PubMed Central

[9] Anwarul S, Dahiya S. Rectified DenseNet169-based automated criminal recognition system for the prediction of crime prone areas using face recognition. J Electron Imaging. 2022;31(4):043055.10.1117/1.JEI.31.4.043055Search in Google Scholar

[10] Machuve D, Nwankwo E, Mduma N, Mbelwa J. Poultry diseases diagnostics models using deep learning. Front Artif Intell. 2022 Aug;5:733345.10.3389/frai.2022.733345Search in Google Scholar PubMed PubMed Central

[11] Sreenivasulu K, Khan HA, Damini K, Akhila M, Bharathi G. Detection of chicken disease based on day-age using pre trained model of CNN. In International Conference on Mining Intelligence and Knowledge Exploration 2023 Jun 28. Cham: Springer Nature Switzerland; p. 13–24.10.1007/978-3-031-44084-7_2Search in Google Scholar

[12] Ducatelle R, Goossens E, De Meyer F, Eeckhaut V, Antonissen G, Haesebrouck F, et al. Biomarkers for monitoring intestinal health in poultry: present status and future perspectives. Vet Res. 2018 Dec;49:1–9.10.1186/s13567-018-0538-6Search in Google Scholar PubMed PubMed Central

[13] Machuve D, Nwankwo E, Lyimo E, Maguo E, Munisi C. Machine learning dataset for poultry diseases diagnostics-PCR annotated (Version 3) [Data set]. Zenodo; Dec 2021. 10.5281/zenodo.5801834.Search in Google Scholar

[14] Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2017. p. 4700–8.10.1109/CVPR.2017.243Search in Google Scholar

[15] He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016. p. 770–8.10.1109/CVPR.2016.90Search in Google Scholar

[16] Hochreiter S. The vanishing gradient problem during learning recurrent neural nets and problem solutions. Int J Uncertainty, Fuzziness Knowl Syst. 1998 Apr;6(2):107–16.10.1142/S0218488598000094Search in Google Scholar

[17] Lin M, Chen Q, Yan S. Network in network. Computing Research Repository (CoRR), arXiv:1312.4400, 2013.Search in Google Scholar

[18] Ioffe S. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167. 2015.Search in Google Scholar

[19] Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res. 2014 Jan;15(1):1929–58.Search in Google Scholar

[20] Garbin C, Zhu X, Marques O. Dropout vs batch normalization: an empirical study of their impact to deep learning. Multimed Tools Appl. 2020 May;79(19):12777–815.10.1007/s11042-019-08453-9Search in Google Scholar

[21] Maas AL, Hannun AY, Ng AY. Rectifier nonlinearities improve neural network acoustic models. Proceedings of the 30th International Conference on Machine Learning, Atlanta, Georgia, USA, 2013.Search in Google Scholar

[22] Smith LN. Cyclical learning rates for training neural networks. In 2017 IEEE winter conference on applications of computer vision (WACV) 2017 Mar 24. IEEE; p. 464–72.10.1109/WACV.2017.58Search in Google Scholar

[23] Shorten C, Khoshgoftaar TM. A survey on image data augmentation for deep learning. J Big Data. 2019 Dec;6(1):1–48.10.1186/s40537-019-0197-0Search in Google Scholar

[24] Afonja T. Accuracy paradox. Towards Data Sci. 2017 Dec;8:39, https://towardsdatascience.com/accuracy-paradox-897a69e2dd9b. [Accessed: Apr. 13, 2024].Search in Google Scholar

[25] Yogi KK, Yadav SP. Chicken diseases detection and classification based on fecal images using EfficientNetB7 model. 2024;11(1):314–30.10.5109/7172288Search in Google Scholar

[26] Degu MZ, Simegn GL. Smartphone based detection and classification of poultry diseases from chicken fecal images using deep learning techniques. Smart Agric Technol. 2023 Aug;4:100221.10.1016/j.atech.2023.100221Search in Google Scholar

[27] Wang J, Shen M, Liu L, Xu Y, Okinda C. Recognition and classification of broiler droppings based on deep convolutional neural network. J Sens. 2019;2019(1):3823515.10.1155/2019/3823515Search in Google Scholar

[28] Aziz NA, Othman MF. Binary classification using SVM for sick and healthy chicken based on chicken’s excrement image. Pertanika J Sci Technol. 2017 Jan;25:315–24.Search in Google Scholar

Received: 2024-10-01

Revised: 2024-11-19

Accepted: 2024-12-10

Published Online: 2025-03-24

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/nleng-2024-0081

Keywords for this article

DeepFowl; disease prediction; chicken excreta; poultry health; deep learning; sustainable development goals

Creative Commons

BY 4.0