Research Of Instance Segmentation Algorithm For Adhesive Objects

In recent years,the theoretical research and related technologies related to deep convolution neural network have developed rapidly.Although the semantic segmentation algorithm and object detection algorithm have achieved good prediction results,the instance segmentation algorithm has encountered various difficulties in the actual research process,such as inaccurate mask edge segmentation,laborious and time-consuming labeling.In addition,the problems of mask "leakage",missed detection and false detection caused by adhesive object have not been well solved.Therefore,this paper aims to study and improve the instance segmentation algorithm in the case of adhesive object.Firstly,this paper introduces the development status of image semantic segmentation algorithm and image instance segmentation algorithm which are based on deep learning,traditional adhesive image segmentation algorithm and attention mechanism.This paper studies the existing representative single-stage instance segmentation algorithm YOLACT algorithm and two-stage instance segmentation algorithm Mask R-CNN algorithm,and introduces the key theoretical parts of these two algorithms.In addition,the adhesive object data set created by depth image in this paper is introduced,including the collection and production of data set,and the evaluation indexes to evaluate the performance of instance segmentation algorithm are also introduced.Configure the running environment of the algorithm and design experiments to analyze the performance of the two algorithms,which provides a direction for the research and improvement of instance segmentation algorithm in the case of adhesive objects.Secondly,the research and improvement are based on the YOLACT single-stage instance segmentation algorithm.As a representative of the single-stage instance segmentation algorithm,YOLACT algorithm will miss detection and false detection of some objects in the face of adhesive objects.Through the design modification of the backbone network,it has stronger multi-scale and feature extraction ability.In addition,the PAFPN feature connection module is built based on PANet,which can further help convolutional neural network obtain effective and rich semantic fusion features.The spatial attention module is constructed based on GCNet and Empirical Attention respectively.The attention module constructed based on GCNet or the attention module constructed based on Empirical Attention is introduced into the residual block,which makes the network learn parameters with clearer directivity and richer expression ability in the process of model training.And then cascade the two constructed attention modules,a more effective cascade spatial attention mechanism is obtained.In order to optimize the inaccurate location of the bounding box,this paper optimizes the regression loss function to improve the accuracy of the prediction of the location of the bounding box.Finally,the two-stage instance segmentation algorithm based on Mask R-CNN is studied and improved.As a representative of the two-stage instance segmentation algorithm,Mask R-CNN algorithm performs well in prediction accuracy,but there will also be missed detection and false detection of some objects in the face of adhesive objects.The object detection network which is the part of the instance segmentation algorithm is designed and modified,and the training effects of the instance segmentation model based on HRNet,Cascade R-CNN and Swin object detection network are compared.The feature connection module is studied and designed,and the built up sampling operation is used to improve the feature connection.By introducing deformable convolution into the backbone network,the ability of the model to learn the invariance of complex objects in the training process is enhanced.After experimental comparison,the best position to introduce deformable convolution into the network is selected.By designing and improving the original network structure,the average accuracy of the bounding box and mask has been improved,which shows that the problems of missing detection,false detection and poor mask quality caused by adhesive objects have been solved to a certain extent.