Research On Pig Identification Based On Mask-RCNN In Deep Learning

With the arrival of the era of intelligence,artificial intelligence technology has become very popular,especially image recognition technology,which is widely used in medicine,military,education and other major fields,but it is slightly insufficient in aquaculture.Traditional pig breeding relies on manual patrol and observation to ensure the health of pigs.Current pig detection technologies are based on traditional image processing methods,which not only have low accuracy,but also consume a lot of time and cost.In order to solve such problems,it is necessary to adopt pig detection technology based on deep learning,which makes the detection faster and more accurate.In order to improve the accuracy of the position location of pig individual detection,this paper proposes a pig individual detection technology based on deep learning.This framework improves the loss function calculation method of the boundary frame of position location on the traditional Mask-RCNN(Mask Region Convolutional Neural Networks)model to ensure that the model is positioned more accurately.Mainly divided into the following three research contents,(1)Aiming at the problem of insufficient training datasets,self-made datasets are adopted,which include the following three methods: first,manual dataset labeling is carried out by using labelme datasets production software.Second,the diversity of extended datasets are automatically generated from images.Thirdly,data enhancement technology is used to cut,rotate,mirror and other image transformation methods for labeled pictures to increase the richness of datasets.(2)Aiming at the problem of slow detection speed of traditional detection methods,deep learning method is introduced and using Mask-RCNN network model to training images.Due to the special RPN(Region Proposal Network)layer in the model,the input size of the image can be any size,and the multi-size formed by the transformation of the images in the self-made datasets are also solved.The experiment shows that the detection speed of this method is greatly improved compared with the traditional image detection method.(3)Aiming at the problem of low recognition accuracy,the Mask-RCNN model is improved,and a loss function based on KL(Kullback-Leibler)divergence is proposed to replace the traditional box regression loss function and a soft-NMS(soft non-maximum suppression)method is proposed to replace the NMS(non-maximum suppression).The method quantifies the difference between the real error and the standard error by KL divergence,minimizes the difference between the two as a loss function,and replaces the direct zeroing operation by reducing the confidence of the box with IOU(Intersection over Union)value larger than the threshold value in NMS.Compared with the traditional model,the method based on KL divergence and soft-NMS has the advantages of more accurate location.This paper builds an improved Mask-RCNN model,The self-made pigdatasets are used to train the model to ensure the feasibility of the experiment.Using four comparative experiments it is concluded that under the same training parameters and the same network model,compare with the traditional method,the improved method improves the accuracy by more than 2% and has a more accurate location.Compared with previous researchers' Mask-RCNN experiments on individual pig detection,this algorithm improves the accuracy of individual pig location by 1.8%.Secondly,experiments are carried out on different numbers of pig data sets,and it is concluded that the improved Mask-RCNN network model has higher recognition accuracy when the data sets are more abundant.Finally,the comparative experiments of fast-RCNN(fast Region Convolutional Neural Network)and faster-RCNN(faster Region Convolutional Neural Network)based on MS COCO(Microsoft Common Objects in Context)data sets are carried out respectively,and the results show that the improved model improves the accuracy of position location by 1.6% and2.5%.In the comparative experiment based on COCO data set with previous fast-RCNN and faster-RCNN,the algorithm used in this paper has improved the average accuracy of regression box by 1.5% and 0.8% respectively.