Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis framework

The widespread occurrence of non-laying hens in commercial poultry flocks significantly affects the management and efficiency of egg production. The current identification of non-laying hens in commercial layer operations relies predominantly on manual empirical methods involving visual inspection a...

Full description

Saved in:
Bibliographic Details
Main Authors: Wenxiang Qin, Xiao Yang, Liyang Yu, Yujia Chen, Junjie Wang, Yan Zhou, Weichao Zheng
Format: Article
Language:English
Published: Elsevier 2025-12-01
Series:Smart Agricultural Technology
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2772375525004150
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849316898472198144
author Wenxiang Qin
Xiao Yang
Liyang Yu
Yujia Chen
Junjie Wang
Yan Zhou
Weichao Zheng
author_facet Wenxiang Qin
Xiao Yang
Liyang Yu
Yujia Chen
Junjie Wang
Yan Zhou
Weichao Zheng
author_sort Wenxiang Qin
collection DOAJ
description The widespread occurrence of non-laying hens in commercial poultry flocks significantly affects the management and efficiency of egg production. The current identification of non-laying hens in commercial layer operations relies predominantly on manual empirical methods involving visual inspection and physical touch. These labor-intensive and time-consuming processes not only lack standardized protocols for objective assessment but also induce physiological stress in poultry populations, thereby contradicting contemporary animal welfare standards. This study proposes an intelligent detection framework for non-laying hens by leveraging a proprietary dataset comprising 977 images of chickens that capture multi-perspective views and behavioral patterns. We systematically compared the feature extraction capabilities of five CNN architectures (ResNet50, ResNeXt50, ResNeXt101, EfficientNet, and ConvNeXt) for non-laying hen identification, analyzed the performance enhancement effects of channel attention and self-attention mechanisms, employed Grad-CAM visualizations to interpret deep feature representations, systematically investigated the contribution of distinct regions to the recognition results, and ultimately established an objective classification framework for non-laying hens. Experimental results demonstrated that ConvNeXt achieved superior classification performance for non-laying hens, with accuracy, precision, recall, and F1 scores of 94.81 %, 93.48 %, 97.37 %, and 95.16 %, respectively. The integration of squeeze-and-excitation (SE) attention mechanisms further enhanced ConvNeXt's classification metrics with accuracy, precision, recall, and F1 scores of 1.29 %, 7.12 %, 2.37 %, and 4.68 %. Although Vision Transformers (ViTs) and Swin Transformers exhibit constrained performances due to dataset scale limitations, they demonstrate potential recognition capabilities. Heatmap visualization analysis revealed that the head features contributed most significantly to identification, followed sequentially by the feet, tail, wings, abdomen, and back regions, suggesting that feather distribution patterns and phenotypic characteristics in these areas may serve as critical biomarkers for non-laying hen classification.
format Article
id doaj-art-ff23876eae93453eb126e5fcf2734cad
institution Kabale University
issn 2772-3755
language English
publishDate 2025-12-01
publisher Elsevier
record_format Article
series Smart Agricultural Technology
spelling doaj-art-ff23876eae93453eb126e5fcf2734cad2025-08-20T03:51:25ZengElsevierSmart Agricultural Technology2772-37552025-12-011210118410.1016/j.atech.2025.101184Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis frameworkWenxiang Qin0Xiao Yang1Liyang Yu2Yujia Chen3Junjie Wang4Yan Zhou5Weichao Zheng6Department of Agricultural Structure and Environmental Engineering, College of Water Resources and Civil Engineering, China Agricultural University, 100083, Beijing, ChinaDepartment of Agricultural Structure and Environmental Engineering, College of Water Resources and Civil Engineering, China Agricultural University, 100083, Beijing, China; Key Laboratory of Agricultural Engineering in Structure and Environment, Ministry of Agriculture and Rural Affairs, 100083, Beijing, China; Beijing Engineering Research Center on Animal Healthy Environment, 100083, Beijing, ChinaFuzhou Magilan Intelligent Technology Co., Ltd., 350011, Fuzhou, ChinaDepartment of Agricultural Structure and Environmental Engineering, College of Water Resources and Civil Engineering, China Agricultural University, 100083, Beijing, ChinaDepartment of Agricultural Structure and Environmental Engineering, College of Water Resources and Civil Engineering, China Agricultural University, 100083, Beijing, ChinaZhucheng Zhongyu Electrical and Mechanical Equipment Co., Ltd., 262200, Zhucheng, ChinaDepartment of Agricultural Structure and Environmental Engineering, College of Water Resources and Civil Engineering, China Agricultural University, 100083, Beijing, China; Key Laboratory of Agricultural Engineering in Structure and Environment, Ministry of Agriculture and Rural Affairs, 100083, Beijing, China; Beijing Engineering Research Center on Animal Healthy Environment, 100083, Beijing, China; Corresponding author.The widespread occurrence of non-laying hens in commercial poultry flocks significantly affects the management and efficiency of egg production. The current identification of non-laying hens in commercial layer operations relies predominantly on manual empirical methods involving visual inspection and physical touch. These labor-intensive and time-consuming processes not only lack standardized protocols for objective assessment but also induce physiological stress in poultry populations, thereby contradicting contemporary animal welfare standards. This study proposes an intelligent detection framework for non-laying hens by leveraging a proprietary dataset comprising 977 images of chickens that capture multi-perspective views and behavioral patterns. We systematically compared the feature extraction capabilities of five CNN architectures (ResNet50, ResNeXt50, ResNeXt101, EfficientNet, and ConvNeXt) for non-laying hen identification, analyzed the performance enhancement effects of channel attention and self-attention mechanisms, employed Grad-CAM visualizations to interpret deep feature representations, systematically investigated the contribution of distinct regions to the recognition results, and ultimately established an objective classification framework for non-laying hens. Experimental results demonstrated that ConvNeXt achieved superior classification performance for non-laying hens, with accuracy, precision, recall, and F1 scores of 94.81 %, 93.48 %, 97.37 %, and 95.16 %, respectively. The integration of squeeze-and-excitation (SE) attention mechanisms further enhanced ConvNeXt's classification metrics with accuracy, precision, recall, and F1 scores of 1.29 %, 7.12 %, 2.37 %, and 4.68 %. Although Vision Transformers (ViTs) and Swin Transformers exhibit constrained performances due to dataset scale limitations, they demonstrate potential recognition capabilities. Heatmap visualization analysis revealed that the head features contributed most significantly to identification, followed sequentially by the feet, tail, wings, abdomen, and back regions, suggesting that feather distribution patterns and phenotypic characteristics in these areas may serve as critical biomarkers for non-laying hen classification.http://www.sciencedirect.com/science/article/pii/S2772375525004150Non-laying hensPhenotypic characteristics analysisDeep learningImage classificationAttention mechanism
spellingShingle Wenxiang Qin
Xiao Yang
Liyang Yu
Yujia Chen
Junjie Wang
Yan Zhou
Weichao Zheng
Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis framework
Smart Agricultural Technology
Non-laying hens
Phenotypic characteristics analysis
Deep learning
Image classification
Attention mechanism
title Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis framework
title_full Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis framework
title_fullStr Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis framework
title_full_unstemmed Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis framework
title_short Automated non-laying hen identification via deep learning: A phenotypic characteristics attribution analysis framework
title_sort automated non laying hen identification via deep learning a phenotypic characteristics attribution analysis framework
topic Non-laying hens
Phenotypic characteristics analysis
Deep learning
Image classification
Attention mechanism
url http://www.sciencedirect.com/science/article/pii/S2772375525004150
work_keys_str_mv AT wenxiangqin automatednonlayinghenidentificationviadeeplearningaphenotypiccharacteristicsattributionanalysisframework
AT xiaoyang automatednonlayinghenidentificationviadeeplearningaphenotypiccharacteristicsattributionanalysisframework
AT liyangyu automatednonlayinghenidentificationviadeeplearningaphenotypiccharacteristicsattributionanalysisframework
AT yujiachen automatednonlayinghenidentificationviadeeplearningaphenotypiccharacteristicsattributionanalysisframework
AT junjiewang automatednonlayinghenidentificationviadeeplearningaphenotypiccharacteristicsattributionanalysisframework
AT yanzhou automatednonlayinghenidentificationviadeeplearningaphenotypiccharacteristicsattributionanalysisframework
AT weichaozheng automatednonlayinghenidentificationviadeeplearningaphenotypiccharacteristicsattributionanalysisframework