Research On The Identification Method Of Feeding Pigs Based On Deep Learning

Pioneering,innovative,modern and intelligent animal husbandry industry is an important means to realize large-scale and precise pig breeding.In the face of large-scale breeding environment,traditional manual observation methods are time-consuming and laborious,and precision breeding is inefficient;computer vision is an effective auxiliary technology for information processing,provides an automated,low-cost,and stress-free identification method.In the breeding farm,the health status of pigs can be effectively grasped through the identification of pigs and the monitoring of their feeding patterns,and the efficiency of precision breeding can be improved.This paper takes group-raising pigs as the research object and uses deep learning methods to identify the identity of feeding pigs.The research content and conclusions are as follows:First,this article uses the method of target detection and pig body structure model to identify the feeding behavior of pigs.Train the Tiny-Yolo v3 target detection network through the self-built data set to recognize the whole pig,head and tail respectively.The network model trained in the experiment has an average detection accuracy of 99.38% on the test set,and its detection rate reaches 48 frames per second,which has good real-time performance.Use the trained model to identify and locate only the target in the feeding area,and combine the pig body structure model to associate the pig body with the head and tail to determine the position of each pig’s head.On this basis,a static state is realized.The algorithm on the picture based on the feeding area occupancy rate predicts the feeding behavior,and calculates the duration of this behavior in a 2-second video segment to identify whether the feeding behavior occurs.The accuracy rate of this algorithm for pigs’ feeding behavior recognition is up to 95.72%.Secondly,the identification of pigs needs to be studied on individual pigs.In this paper,we use Full Convolutional Neural Network(FCN)and Canny edge detection methods to segment individual pigs and extract foreground targets.The Tiny-Yolo v3 target detection network is used to extract complete images of individual pigs,and FCN is used to roughly segment the extracted images.In order to obtain more accurate segmentation results,filter algorithm and Canny operator are used to denoise and edge detection of the original image,and then the pig contour of the rough segmentation result is converged to the contour after edge detection,and the segmentation accuracy of the image on the test set Reached 93.51%.Finally,for the self-built individual pig identification database,this paper uses an improved convolutional neural network(CNN)to train and identify individual pig images.By comparing the pig identification effects of Alex Net and Goog Le Net(Inception v1),a simpler Alex Net model is selected and its internal structure is improved: batch normalization is used instead of local response normalization,and the number of neurons in the fully connected layer is reduced.Obtained a lighter CNN model(Tiny-Alex Net,TA).After the improvement,the amount of parameters is reduced by 87.6%,and the recognition accuracy is increased from 90.06% to94.13%.The performance of the model is effectively improved,and it is compatible with the pig feeding behavior recognition algorithm in combination,the trained network is used to identify the identity of the feeding pigs,and the feeding time of each pig is counted.The accuracy of the identification of the feeding time of individual pigs is up to 97.16%.