Research On Automatic Recognition And Counting Methods Of Caged Hens And Eggs Based On Deep Learning

Breeding scale and production capacity of Chinese layer hens rank first in the world,with egg production of about 28,627,000 tons by the end of 2021.Currently,three-dimensional cage farming is main model of egg farming,the efficient inspection and dynamic clearing of lowlaying hens in this model has become an important demand for the egg farming industry.However,the high density of cage breeding and the uneven distribution of light between hens make the accurate detection and counting of hens and eggs a challenge.To address these problems,in this study a lightweight network YOLOv7-tiny-DO based on YOLOv7-tiny for hens and eggs detection in cage mode was proposed,and an automated counting method in different cages was designed.Then,the optimal model was deployed to edge computing devices to detect hens and eggs and to count in different cages in real time.Finally,Daily egg production information were obtained,a low-laying cage discrimination algorithm based on Adaboost was constructed,and a cage information monitoring system for low-laying hens in cage mode was developed.The main work of this study is as follows:(1)A hens and eggs image dataset was constructed.In this study,an automated data acquisition platform was built for the hen coop structure,and the JRWT1412 camera was used to capture hens and eggs video data,from which a total of 2146 caged hens and eggs images were obtained for the construction of the dataset.(2)A YOLOv7-tiny-DO based caged hens and eggs detection model was proposed.In this thesis,the YOLOv7-tiny network structure was improved by applying ELU activation function,depthwise over-parameterized depthwise convolutional layer and coordinate attention mechanism.Test results showed that the average precisions(APs)of YOLOv7-tiny-DO in identifying hens and eggs were 96.9% and 99.3% respectively,which were 3.2 and 1.4percentage points higher than those of YOLOv7-tiny for hens and eggs respectively.(3)An automated cage split counting algorithm for hens and eggs in cage mode was designed.The hens and eggs of counting in different cages was designed for the layout of hen cages in the hen house and calibrated using a monocular camera calibration model.30 hen cages were selected for counting tests in a real-world scenario for 3 days.The results showed that the average accuracy of hens and eggs counting for the three test batches was 96.8% and 96.4%respectively,with an average absolute error of 0.12 hens and 0.09 eggs per cage respectively.(4)An Adaboost-based cage discrimination model for low laying hens in cage culture was proposed.The information of daily hens and eggs was extracted,based on which the daily egg production rate was calculated,and a multi-dimensional feature vector was constructed to build a cage discrimination model for low laying hens under cage rearing mode.The results showed that the classification accuracy of Adaboost was up to 98.8%.(5)A cage monitoring system for low yielding hen cages in cage mode was developed.In this paper,Py Qt5 on Ubuntu 18.04 system was used to design the system interface,My SQL database was applied to store hens and eggs information,and Python language was used for integrating three algorithm modules of target detection,splitting cage counting,and low-laying cage discrimination into the system.The system’s simple interface allowed farm workers to identify low-laying cages by simply checking the abnormal data interface on a daily basis,thus quickly identifying low-laying hens.In this study,data and technical support for digital hens house can be provided.