Detection Method For Behaviors Of Sows Based On Deep Learning

China is the largest producer and consumer of pork around the world.However,intermittent observations by breeders is highly involved in traditional pig industry household to obtain individual information of pigs,for the sake of determining pigs in abnormal conditions.This lays huge workload on breeders,along with possible risks of low efficiency,poor effectiveness and so on.By monitoring the behavior of the sows through machine vision,it is possible to provide early warning,which can effectively protect the health of sows and piglets,thus improve production efficiency.Traditional machine vision methods for monitoring sows behavior usually refer to one single scene,with the defect of poor robustness.To this end,this paper developed methods based on deep learning to monitor behaviors of pregnant sows.The main research contents and conclusions are listed below:(1)A head and tail recognition algorithm was proposed based on image processing.Individual pig was first identified using DeepLab model,then three contour descriptors of chain code,complex number and polar coordinates,in combination of three typical classifiers of Fisher linear discriminant,K-nearest neighbor(KNN)and Naive Bayes,were utilized and compared in head and tail recognition.It is found that when combination of complex description and KNN was used,the pig head and tail recognition effect was the best,and accuracy rate could reach 85%.(2)Three kinds of fine segmentation algorithms for improving the accuracy of segmentation images of deep learning models were proposed.Aiming at the problem that the deep learning model cannot further improve the segmentation accuracy,the initial segmentation graph obtained by the deep learning model,combining with its edge,was improved by segmentation algorithms based on the watershed algorithm,the region growing algorithm,and Canny edge detection algorithm,respectively.The algorithm's fine segmentation algorithm could improve the image detection accuracy of specific scenes by 0.73%,4.41%and 4.98%.(3)A method for recognizing pig posture using a single image was proposed.Using the established pig recognition model and the pig head and tail recognition algorithm,the pig image to be identified was adjusted to the horizontal state,and the pig was on the same side,thus establishing the LeNet model of pig posture recognition,and recognition accuracy was about 95%.Using this model to study behavior of sows before delivery,results showed that sows would significantly increase the amount of activity and expand the range of activities before delivery.(4)An image preprocessing method based on neighborhood search method for recovering monitoring watermark was improved.Gaussian difference optimized watermark localization was added for non-transparent watermarking.A watermark localization method based on neighborhood pixel difference was proposed for transparent watermarking,which could eliminate watermark and restore image effectively.