Research On Chicken Part Detection And Recognition Technology Based On Deep Learning

At present,in the poultry processing industry,manual sorting of poultry parts(chicken legs,wings and wing roots)is usually used.This manual sorting method has the advantages of low processing efficiency,time-consuming,labor-consuming and high cost,which is difficult to meet the needs of modern production.With the rapid development of machine vision technology,machine vision technology is widely used in various production and processing fields,especially in the field of image recognition.Using deep learning technology to detect bird parts and improve the detection speed is of great significance to speed up the modernization process of poultry processing industry.This paper takes the chicken part as the research object,carries out the image acquisition,image segmentation and image recognition of the chicken part,and focuses on the chicken part detection and recognition technology based on deep learning.In this paper,the image acquisition system of chicken part is built,and the training set,verification set and test set of chicken part image are obtained.These data are used for chicken part detection and classification.The main research contents of this paper are as follows:(1)The image acquisition system is built.According to the requirements of poultry part detection system,the appropriate types of industrial camera,lens,light source and PC in the image acquisition system are selected.Under the condition of ensuring automatic focusing,1000 chicken part images are collected,and then the data is enhanced and expanded to 6000.The chicken part data set is established.At the same time,labelmg software is used for manual labeling,and finally the image is processed to make the image clearer and more recognizable.(2)Aiming at the problem of chicken position target recognition,a detection method based on chicken position deep learning is proposed.The characteristics of four target detection algorithms of convolutional neural network and Yolo,fast r-cnn,SSD and yoov3 network models are analyzed.The trained models of Yolo,fast r-cnn,SSD and yoov3 are tested respectively.The data show that the detection map values are 89.1%,88.7%,89.3% and 92.1% respectively,and the single detection time is0.062 s,0.069 s and 0.049 s respectively,0.060 s,the results show that the effect of Yolo V3 detection model is good.(3)An improved AM-CIOU-YOLO V3 target detection algorithm is proposed.Two attention mechanisms,STN and SE,were added to the main network of Yo Lo V3,and IOU in Yo Lo V3 was replaced with CIOU to obtain the final improved model amCiou-YOLO V3,which was compared with Yo Lo V3,Faster RCNN and SSD algorithms.The results show that the F value of the improved method can reach97.3%,and the detection time of single sheet is 0.026 s.(4)Aiming at the problem of target detection of chicken parts,this paper puts forward an algorithm based on semantic segmentation,analyzes the semantic segmentation of FCN and If supervision,and conducts experiments on three trained models of U-NET,FCN and Seg Net.Through experimental comparison,the U-NET semantic segmentation algorithm with the best accuracy and the shortest time is selected.Its accuracy and average crossover ratio are 93.63 %and 85.37,and its time consumption is 156 ms.In conclusion,the detection and recognition technology scheme of chicken parts is proposed in this paper,which can achieve better recognition effect and satisfy realtime sample processing.Two-stage classification and recognition of chicken parts can further improve the recognition accuracy of chicken parts and improve the level of intelligent recognition for poultry processing.