The Method Of Sheep Trunk's Segmentation And Mutipart Detection Based On Machine Learning

The segmentation of sheep trunk is the most difficult part in the process of sheep slaughtering and processing.In some developed countries,automatic slaughter and division processing has been realized,but in China it has been completed by manual with semi-automatic machinery.Because there is no half-splitting process in domestic mutton segmentation,foreign mutton segmentation equipment and technology can not be fully applicable to the national conditions of our country.Therefore,it is of great significance to study the method of automatic segmentation and identification of sheep trunk which can improve the automatic processing level of sheep skeleton segmentation and improve the food safety of mutton products.In this study,according to the actual demand in the process of sheep skeleton segmentation in domestic enterprises,in order to ensure that the robot system can accurately obtain the position of sheep trunk segmentation,the location of sheep trunk segmentation is predicted by machine vision and machine learning methods,and the detection and recognition of split mutton are deeply studied.The main research contents and conclusions are as follows:(1)Image segmentation and feature extraction of sheep skeleton.In view of the complex biological characteristics of sheep trunk individuals,the study of the shape and position distribution characteristics of sheep trunk in the image is beneficial to achieve the segmentation of sheep skeleton.The coordinate parameters of the minimum-outer-rectangle is an important representation of the shape and position of the trunk and chest of the sheep skeleton,which can be used to describe the shape of the sheep.The characteristic parameters of sheep trunk are calculated by grayscale,binarization,morphological operation and contour recognition.A recognition method is proposed for the difficult recognition of lumbar and neck,which is based on machine vision and cart decision tree.Color feature and spatial position feature are important features of lumbar and neck,which can be used for segmentation,recognition and feature parameter extraction of lumbar and neck.According to the images of lumbar and neck in RGB,HSV and YCb Cr color space,the color is extracted by using the distribution of green in HSV color space,combining the extracted results with YCb Cr images to get the lumbar spine,neck and interference factors.According to the central coordinates and connected area of the lumbar,neck and interference factors,a cart decision tree classification method is established to identify the lumbar and neck and extract the characteristic parameters.A total of 24 sets of shape and position feature coordinate parameters were extracted from four parts of 1000 groups of images,which provides data base for building the model of machine learning.(2)Sheep's feature segmentation based on improved Deep Lab V3+.Study the infrastructure of deeplabv3+,and adjust the Res Net.The ASSP module were improved by changing the expansion rate of hole convolution and introducing deformable convolution.Eight kinds of deep labv3+deformation networks were built.The MIo U,PA and F values of improved network 6 are 0.849,0.870 and 0.879,which are1.6,2.2 and 0.7 percentage points higher than the original deeplabv3+.(3)Position prediction of sheep trunk segmentation based on machine learning.Studying the application of machine learning method in segmentation position prediction is conducive to the realization of sheep trunk segmentation with different size and quality.Feature engineering is an important part of machine learning.In this study,correlation test,outlier detection,feature construction,data translation and normalization are performed on the feature parameter set of sheep skeleton.Designed and developed Lasso,Ridge,SVR and GBDT model.The evaluation of the model is based on three indexes,which are R~2,MSE and absolute mean of residual.Lasso and SVR models have the best prediction effect,and Ridge and GBDT models have the second.The integrated learning algorithm is designed and developed based on the Bayesian optimization method.When Lasso,SVR and GBDT are integrated in the ratio of 0.30:0.25:0.45,the model effect is the best.The fitting degree R~2 is 0.947,the mean square error MSE is 8.26,and the absolute average residual error is 2.0 pixels.(4)Detection of mutton segmentation based on deep convolution neural network.In view of the classification problem of segmented mutton,research on the detection and recognition technology of segmented mutton is helpful to realize automation of mutton classification process.Res Net-30,38,44,56,89 and 113 classification networks were built based in Res Net.Compared with VGG-16,Res Net-34,50 and 101,Res Net-38 was selected as the feature extraction and classification network of mutton segmentation.The top-1 error was 0.253%.RPN network layer,ROI pooling layer and resnet-38 are introduced to design a detection and recognition network which named Sheep D-Net.The improved model was validated on the test set,with an average accuracy of 84.5%and a processing speed of 0.154s/piece.Compared with YOLOV3,Faster-RCNN and Couple Net,the average accuracy of map increased by 7.0,6.3 and 0.9 percentage points.And the recognition speed increased by 0.071s compared with Couple Net.(5)The segmentation test ofsheep trunk.The test platform was built to verify the actual prediction ability of integrated learning.The selection of manipulator,the construction of fixture system,the design and development of trajectory analysis and optimization algorithm,and the development of visual monitoring interface are completed.According to the quality,the samples were divided into three groups.The number of samples were 2,4 and 2.The average deviation distance was 3.45mm.