Weakly Supervised Object Detection Method And Its Application

As a long-standing basic problem,object localization and detection is the key research content in the field of new generation computer vision.The research purpose of object localization and detection is to quickly and accurately judge whether there are object instances of a given category from a given image,and return the localization information of its coverage to the object instances in the image.When only image level labels are used as supervision information,the object localization model based on weak supervision learning often activates only the part object and ignores the whole object region when optimizing image classification,and expanding the active part to the whole region will reduce the ability of model classification.Based on the core problem of"how to improve the performance of model positioning the whole object region on the basis of ensuring the good performance of image classification",this thesis puts forward a new idea of object positioning based on the joint training of foreground activation and background suppression.The method of object localization is used in the object detection task,starting from the practical problem,to solves the challenges faced by traffic sign object detection in actual scenes,and makes an in-depth study on the practical application of object localization and detection in life.The main contents of this thesis are as follows:(1)In this thesis,a weakly supervised obj ect localization method for joint training of activating object foreground and inhibiting activating background is designed,which makes the network model effectively distinguish the difference between foreground object and background environment.The method proposed uses the different semantic information in the feature map to expand the feature activation according to the level of the category of object,and then activates the object divergently level by level,and finally obtains the prediction of the foreground activation map.Using foreground guidance,a background activation suppression module is designed,which is combined with foreground activation prediction to obtain the final object localization result.Experiments show that on the CUB-200-2011 and ImageNet datasets,the network model proposed can effectively locate the whole object region and improve the accuracy of obj ect localization.(2)For the specific task of object detection,which is traffic sign object detection,this thesis optimizes the existing two-stage traffic sign object detection network.For the detection task in the actual scene,the size of the traffic signs to be detected in the image is random,which leads to the unsatisfactory effect of the baseline model for locating small-size traffic signs.In this thesis,following the idea of object localization,the idea of multi-scale object localization and detection is introduced,and an image segmentation method is proposed to "enlarge" these small-scale objects.Using the ability of object localization in location recognition,the network is optimized to locate near the traffic sign region,which provides strong support for the detection branch.At the same time,DIoU loss function is used to improve the regression accuracy of anchor frame,so as to improve the accuracy of object detection.The experimental demonstration is carried out on the traffic sign data set of actual scene,and the results show that our method has excellent positioning and detection performance in the actual scene.