Research On Weakly Supervised Instance Segmentation Technology Based On Peak Response Map

The purpose of image instance segmentation task is to identify and detect the individual in the image at the pixel level.It has a wide application prospect in the fields of autopilot,UAV flight,geographic information system,medical aided diagnosis,intelligent robot and so on.In recent years,as deep learning has been widely used in various tasks of computer vision,such as image recognition,object detection,semantic segmentation and other tasks,and has achieved brilliant success,the instance segmentation task has also turned its attention on deep learning and made remarkable progress.Since the training data of the fully supervised instance segmentation model needs a large amount of pixel-level annotations which are extremely costly,the instance segmentation based on weak annotation has gradually attracted extensive attention of researchers.In this context,this paper makes further study on the weakly supervised instance segmentation based on deep learning technology.At present,in many weakly supervised instance segmentation methods,all of them are based on Pascal VOC training data with only image level labels to achieve the training of segmentation network.Due to the lack of real annotation at instance level and pixel level,the accuracy of instance classification,recall rate of detection and integrity of segmentation in the image are difficulties in this task.On this premise,the performance of the model cannot be taken into account.In order to make up for the deficiency that the image class labels can not provide the position information and shape information of the instance effectively,the method based on the peak response map mining and filtering is proposed.Then the classification,detection and segmentation of the instance segmentation task can be effectively solved to improve the overall accuracy of the task.Simultaneously,the concept of multi-scale peak response map is proposed,and a cascade segmentation subnetwork is constructed based on it,which can not only improve the accuracy of the model,but also significantly improve the performance.The main contributions of this paper are as follows:1.Aiming at improving the precision of weakly supervised instance segmentation,an adversarial strategy based on peak response map mining and filtering is proposed.The ultimate goal of the method is to provide pixel-level labels with higher precision for training the instance segmentation algorithms in fully supervised.Before this,the erase strategy was introduced to expand the semantic response region of the objects in the image,and the peak response map algorithm(PRM)was used to perform the peak stimulation and response.Then cluster analysis was carried out based on the deep feature of the peaks,and different peak response maps from the same instance were merged to improve the integrity and diversity of its.On this basis,this method adopts the idea of filtering to design a scoring mechanism based on the peak response map,and iteratively retrieves the best items from the proposal gallery provided by MCG as the mask.Then,the category and shape information were combined to update the confidence to filter the masks,so as to obtain a cleaner,more accurate and complete pixel-level annotations.The proposed method can make up for the defect of rough segmentation caused by insufficient label information,and improve the detection and segmentation accuracy of instance segmentation task under weak supervision.2.Aiming at improving the precision and efficiency of weakly supervised instance segmentation,a cascade segmentation subnetwork based on peak response map is proposed.The multi-scale peak response map is introduced by using the activated peaks from the feature maps of different layers in the backbone network,which can improve the diversity and integrity of the peak response map.In order to achieve more complete and accurate instance segmentation,the method uses instance activate map algorithm(IAM)to fill the incomplete and local peak response map to the complete object with the pseudo pixel level supervision of segmentation mask.Then a cascade segmentation subnetwork is proposed based on it,step by step using different degree of complete peak response map and the segmentation output as the next level network's input for supervised training,and according to the last level segmentation mask to extract more abundant object features for feature enhancement.Therefore,it is no longer necessary for the model to retrieve from the proposal mask gallery in inference.In addition to further refining the quality of segmentation,the speed of segmentation can also be significantly improved.Based on the above strategies,our methods achieve good results on the PASCAL VOC 2012 dataset,which showed a great improvement in accuracy and performance compared with the existing methods,and the effectiveness of each strategy was verified through comparative experiments.