Research On Instance Segmentation Algorithm Of Occluded Pigs Based On Deep Learning

Image instance segmentation is an important research direction of deep learning and artificial intelligence in the field of image processing.It can segment the object at the pixel level on the basis of accurately detecting the object in the image.At present,the research results of image instance segmentation have been widely used in automation industry,geographic information monitoring,intelligent breeding and other fields,but in herd pig farm,the existence of occluded object affects the performance of the algorithm.The instance segmentation algorithm will have two problems in the face of occluded object: one is that the performance of the detector is affected due to the lack of feature information of occluded object,resulting in missed detection of object.Second,the occluded object are affected by the surrounding object and are easy to generate low-quality masks.Based on the research of multi-stage cascaded instance segmentation algorithm HTC,this paper decomposes the instance segmentation framework into object detection network and mask generation network,and improves the two networks respectively to solve the problem of poor effect of occlusion object detection and segmentation in the task of image case segmentation.The main research and work contents are as follows:Firstly,based on the cascaded multi-stage instance segmentation algorithm HTC,this paper tests the performance of the algorithm on the self-made pig data set with a large number of occluded object,and analyzes the problems of the algorithm in occluded object detection and segmentation according to the model detection results.Secondly,in order to optimize occluded object detection,an improved object detection network based on HTC is designed to improve the detection and segmentation accuracy of occluded object.Improve the object detection network,and put forward improvement schemes for feature extraction network,RPN network and non maximum suppression module.The backbone network adopts the residual network CA-Res Net that introduces coordinate attention to help the network capture the location and channel information of the object.The feature connection module designs a multi-scale fusion feature pyramid DF-FPN with feedback mechanism to enhance the object feature information fusion extracted by deep network.The region feature generation module designs Cascade-RPN network to improve the quality of the generated object candidate region.The post-processing module designs a postprocessing filtering module Soft-MNMS combined with mask IOU,which suppresses redundant candidate frames and other object candidate frames belonging to the same object through Gaussian attenuation function,so as to reduce the missed detection of occluded object.Finally,in order to improve the mask accuracy of occluded object,a hybrid mask generation network based on HTC is designed.The mask of the whole image is generated from different scale feature maps obtained from the feature extraction network through the bottom module,and then the high-resolution mask of the object is obtained by weighting the attention map generated by the top module according to the regional features,so as to improve the segmentation accuracy of the example segmentation algorithm.