Research On Weakly Supervised Object Location And Detection Under Multi-scale Feature Fusion

With the application of object location and detection technology in daily life more and more widely,the existing technology in the training process requires large-scale datasets with tags,and weakly supervised technology has the advantage of relying only on image-level tags,so it has caused extensive research.Although this technique has made great progress in recent years,it has obvious gap with strong supervised in multi-scale object detection,local optimal problem and multi-instance problem.Therefore,in order to solve the above problems,two algorithms of weakly supervised object location and weakly supervised object detection are proposed respectively,and a weakly supervised object location and detection system is built.The main research work includes:(1)Aiming at the problem that multi-scale object location effect is poor and it is difficult to capture the complete object boundary,a weakly supervised object location algorithm based on multi-field fine analysis is proposed.Firstly,a multifield fine analysis module is designed to obtain high-resolution features of multiscale objects.Secondly,the random feature selection module is used to obtain the combination of random positions of the feature map and obtain the complete location information of the object.Finally,the object information of the shallow class activation map and the deep class activation map is fused to capture the complete object boundary.Tests on three mainstream data sets show that,compared with the existing weakly supervised localization methods,it has certain advantages in solving the problem of poor localization effect and local optimization of multi-scale objects.(2)A weakly supervised object detection algorithm based on object instance mining is proposed to solve the problems of poor multi-scale object location effect,local optimal and multi-instance of the detector.Firstly,the multi-sample pyramid is constructed to aggregate multi-scale context information.Secondly,the selective search technology is combined with the Grad-CAM++ technology to generate higher quality proposals.Finally,the object instance mining module is used to detect all possible object instances in the image.In addition,the object instance reweighting loss function is designed to learn a large part of each object instance,so as to further improve the performance.The test on VOC dataset shows that it has certain advantages compared with the existing weakly supervised object detection methods.(3)A weakly supervised object localization and detection system is designed and implemented.The main functions of the system are input/output module of pictures,display module of pictures,weakly supervised object localization algorithm module,and weakly supervised object detection algorithm.The two weakly supervised algorithms proposed in this paper effectively improve the performance of multi-scale object location and alleviates the poor multi-scale object detection effect,local optimum problem and multi-instance problem.The weakly supervised object location and detection system provides a solution for object location and detection.