Identification Method Of Key Parts Of Sheep Skeleton And Research And Development Of Automatic Cutting System

As an important part of China’s livestock product processing industry and the main meat consumer goods for urban and rural residents,the demand and production of mutton products are increasing day by day.In the process of mutton processing,the sheep skeleton cutting is a process of cutting the key parts of the sheep skeleton into corresponding mutton multi-parts according to industry standards and public dietary preferences.It is an important link to realize the value-added of mutton products.However,at present,China’s mutton processing enterprises generally complete this link in the form of manual assisted assembly line operation,which has a series of problems such as low production efficiency and poor working environment,It restricts the development of sheep industry in China.In order to respond to the call of the development of"meat industry"and improve the automation and intelligence level of the sheep skeleton cutting process,this paper takes the sheep skeleton and its cut mutton multi-parts as the research object focuses on the actual requirements in the processing of mutton products and combines computer vision technology and automatic control technology to carry out in-depth research on the data generation of sheep skeleton image,the semantic segmentation of three parts of sheep skeleton image(spine,chest,and neck),the prediction of sheep neck clavicle cutting distance,and the real-time classification and detection of mutton multi-parts.Based on the above research results,the automatic cutting control system of sheep skeleton is proposed and the cutting test is carried out.The main research contents are as follows:(1)Research on data generation of sheep skeleton image.Given the difficulties in obtaining sheep skeleton samples and the limited amount of image data,the generation countermeasure network is used to generate sheep skeleton images with complete semantic information,to expand the data scale of sheep skeleton images.To ensure the correct and clear expression of the natural features of each part of the generated sheep skeleton image,the resolution of the original small-scale sheep skeleton image is set to512 pixels×512 pixels based on the principle of scale invariance and normalized.Three kinds of high-resolution image generation countermeasure networks,DCGAN,Sin GAN,and Big GAN,are introduced to train the original image of sheep skeleton.The generator G is used to learn the characteristic distribution law of sheep skeleton and the discriminator D is used to determine the effect of generating sheep skeleton image,to improve the authenticity of generating sheep skeleton image.Through comparative analysis of three network generators G＿Loss,discriminator D＿Loss,which change trend of loss in the case of the same number of iterations,and judge the quality of each generation countermeasure network on the generation ability of sheep skeleton image based on the qualitative evaluation index,and select Big GAN as the final sheep skeleton image generation network to realize the effective expansion of the data scale of sheep skeleton image.(2)Real-time semantic segmentation of three parts of sheep skeleton image.The sheep skeleton is composed of three parts:spine,chest and neck.Each part is bonded to each other.The color,texture,and other characteristics are similar,and the distinction is not significant,which is not conducive to the accurate semantic segmentation of the sheep skeleton image.Aiming at the above problems,a real-time semantic segmentation method of three parts of sheep skeleton image based on ICNet is proposed.Based on the original sheep skeleton image data set,this method establishes a combined data set with the generated sheep skeleton images.Based on the migration learning training model,this method uses the strategy of semantic segmentation of low-resolution image and thinning the segmentation results of a high-resolution image to realize the real-time and accurate semantic segmentation of three parts of the sheep skeleton.PA,MIo U,and the average processing time of a single image are selected as the evaluation criteria for the performance of the three parts real-time semantic segmentation model of sheep skeleton image,and the generalization ability of the model is tested by simulating sheep skeleton images with different illumination levels.In addition,four commonly used image semantic segmentation algorithms,U-Net、Deep Lab V3、PSPNet and Fast-SCNN,are introduced to further compare their comprehensive capabilities.The results show that the segmentation accuracy,MIo U,and average processing time of single image of the three parts real-time semantic segmentation model of sheep skeleton image based on ICNet are97.36%,88.10%,and 87 ms respectively,and the comprehensive ability is the best.(3)Acquisition of the cutting distance between sheep neck and clavicle.The individual sizes of diverse sheep skeletons are different,and the cutting distances of sheep neck and clavicle are different.To realize the correct and accurate cutting of sheep necks and clavicles of different shapes of sheep skeletons in the non-contact state,the prediction of sheep necks clavicle cutting distance is carried out based on Gaussian process regression theory and machine vision technology.By analyzing the main parts of the shape change of sheep skeleton,eight groups of body size characteristics of 120 pairs of sheep skeleton were collected,including rib length,rib width,spine length,spine width,total length,total length after neck removal,and the cutting distance of sheep neck and clavicle.After eliminating the correlation analysis between outliers and each body size characteristic,the body size characteristics strongly related to the change of cutting distance of sheep neck and clavicle were selected.The prediction model of sheep neck clavicle cutting distance was obtained based on Gaussian process regression,and compared with four commonly used regression algorithms:RF,GBRT,SVR,and SLR.Finally,the determination coefficient of the prediction model for the cutting distance of sheep neck and clavicle,the R~2、RMSE and MAE reach 0.97,0.30,0.18 and 0.83,0.29and 0.22 respectively,and the prediction accuracy is the highest.On this basis,using the real-time semantic segmentation results of three parts of the sheep skeleton image,further adjust the segmentation threshold according to the color characteristics,combined with image morphological processing,convert the pixel distance of each body scale of the sheep skeleton into its actual physical length,input the prediction model,and obtain the cutting distance results of the sheep neck and clavicle.The predicted cutting distance data of sheep neck and clavicle output by the model are compared with the data of the manual measurement control group.The average absolute percentage error MAPE of the two groups of data are 3.72%and 6.82%respectively,to correctly obtain the cutting distance of sheep neck and clavicle of different sheep skeletons under the condition of no contact.(4)Real-time classification and detection of mutton multi-parts.The mutton multi-parts produced after sheep skeleton cutting need to be classified and detected.Taking sheep lumbar vertebrae,thoracic vertebrae,neck,abdominal rib,scapula,and leg bones as the research objects,a real-time classification and detection method of mutton multi-parts is proposed by using a single-stage target detection algorithm.In the conveyor belt scene of the sheep slaughtering workshop,the multi-split images containing multiple types and multiple mutton are collected,and the multi-split image data set of mutton is established after image expansion and normalization.Transfer learning is introduced to train the mutton multi-parts classification and detection model based on YOLOv3 and SSD respectively.By comparing their m AP and the average processing time of a single image,the real-time classification and detection model of mutton multi-parts based on YOLOv3 is optimized to realize the real-time classification and detection of mutton multi-parts category and position in the image.The model detection speed is optimized by replacing the feature extraction network with a lightweight neural network.On this basis,the robustness,generalization ability and anti-interference ability of the optimized model are tested by using the additional multi-scale data set with significant multi-scale features,the additional illumination data set representing"bright"and"dark"brightness conditions,and the additional occlusion data set representing occlusion conditions.Through three common target detection algorithms:Mask RCNN,Faster-RCNN,and Cascade-RCNN,and replacing the feature extraction networks as Mobile Netv1,Res Net34,and Res Net50,further comparative experiments are carried out with the optimized model.The results show that for the mutton multi-split image verification set,the m AP and single image detection time of the optimized model are 88.05%and 64.7ms respectively,which is0.26%and 32.67%higher than the original model.It has good robustness,strong generalization ability,and anti-interference ability,and the comprehensive detection ability is the best.(5)Design and experimental analysis of automatic cutting system for sheep skeleton.To realize the intelligent and automatic cutting of sheep skeleton,a two-step cutting process of"cutting three parts of sheep skeleton first,and then cutting the remaining sheep chest"was proposed,and the automatic cutting device for processing three parts of sheep skeleton and automatic cutting device for sheep chest were designed to construct an automatic cutting system.By 3D scanning and reconstructing the spatial three-dimensional structure of sheep skeleton,referring to the shape characteristics and cutting requirements of sheep skeleton,the sheep skeleton profiling fixture is designed for clamping operation in the cutting process.The sheep skeleton chain conveyor is constructed to transport the sheep skeleton in the clamped state at a uniform speed,and then the cutting tools are staggered according to the cutting position of the three parts of the sheep skeleton.The automatic and correct cutting of the three parts of the sheep skeleton is completed by using the sheep neck clavicle intelligent cutting system combined with two groups of movable circular saws.Aiming at the automatic cutting of sheep’s thorax,based on the structural characteristics of sheep’s thorax,the profiling inner mold of sheep’s thorax and the pneumatic clamping system of sheep’s thorax are designed.With the help of the two modes of"air blowing"and"air suction"of the pneumatic clamping system,the profiling inner mold achieves the purpose of clamping before cutting and separating after cutting.According to the cutting specification of sheep thorax,a 5-saw cutting table of sheep thorax was constructed to complete the automatic cutting operation of sheep thorax.The cutting performance test of sheep skeleton automatic cutting system was carried out with three evaluation indexes:average cutting deviationθ,cutting efficiency P,and cutting loss rateη.The results show that the above three evaluation indexes reach 8.6mm,99.5pairs/h,and 1.11%respectively.Aiming at realizing the recognition and detection of key parts of sheep skeleton and improving the automation level of sheep skeleton cutting operation,this study researched sheep skeleton image generation,sheep skeleton 3-part recognition,sheep neck clavicle cutting distance prediction,mutton multi-parts classification,and detection method and sheep skeleton automatic cutting system,to provide technical support for the automation operation of sheep carcass slaughtering and processing production line.