Research On Object Detection Method Based On Multi-instance Weakly Supervised Learning

In recent years,the theory and application of object detection in the field of computer vision have received more and more attention.With the advancement of science and technology,more and more products based on object detection technology are now appearing in our lives,providing a lot of convenience to our lives,but object detection consumes a lot of manpower during the training phase.Material resources are used to annotate pictures,and weakly supervised object detection has entered people's field of vision.The current mainstream weakly supervised object detection methods are multi-instance learning and detectors using end-to-end methods,but there are still shortcomings.There is another unavoidable problem in the object detection of example weakly supervised learning,that is,the problem of nonconvexity.This approach based on end-to-end detectors has two disadvantages: one is that it cannot make full use of contextual information to classify proposals;the other is the part of the object that is most likely to be detected instead of the entire object.Therefore,in order to enable the end-to-end detector to effectively use the context information to classify proposals,in view of the non-convexity problem of the detection result that can only detect the object and the weakly supervised object detection,the main improvements in this paper are as follows:(1)Proposal integrity scoring network(SSNet)is proposed for the inability to effectively use context information to classify proposals.This paper uses average pixels to fill the proposal area in the image and then puts the entire image into SSNet.The more regions contained in the image processed by the object,the higher the output network score.In this paper,by judging the completeness and compactness of proposals based on SSNet output scores,different proposals can be effectively classified.In addition,for the problem that the part of the object may be detected instead of the entire object,this paper introduces the regression box module into the PCL of the existing weakly supervised object detection network.The function of this module is to modify proposals that are different from the basic facts.Moreover,the regression box module is introduced into the SSNet proposed in this paper to make it more accurate.Finally,the effectiveness of this method is verified through multiple sets of comparative experiments.(2)Aiming at the problem of non-convexity in the object detection of multi-instance weakly supervised learning,this paper proposes an online proposal generation(OAG)algorithm to alleviate the problem of non-convexity.During the training process,the performance of the training network will be affected.The difference.This affects the quality of online PGT in this training network.Proposals based on low-quality PGT sampling may delete many positive proposals and disrupt the training process.To solve this problem,a dynamic proposal constraint(DPC)is proposed.DPC can determine different proposal sampling strategies to the current training state.This article defines the training state ?.By adjusting the ? value,find the optimal ? value.Finally,the effectiveness of this method is verified through experiments.