Research On Weakly Supervised Object Detection Algorithm Based On Multiple Instance Learning

With the development of deep convolutional neural network,fully supervised object detection algorithms have made great progress in terms of accuracy and speed.However,the acquisition of detection model with high performance relies on a large numbers of finely labeled training images.For each object,the class and the bounding box of the object need to be marked.The annotation of large-scale training data is expensive and time-consuming.In order to reduce the cost of data annotation,researchers in great numbers have begun to explore weakly supervised object detection algorithms,which only require image-level class annotations.Compared with fully supervised object detection algorithm,weakly supervised object detection algorithm still has a large gap in accuracy and speed.In order to improve the performance of the weakly supervised object detection algorithm,we improve the weakly supervised object detection algorithm from the following three aspects,Firstly,due to the lack of object bounding box annotation,for the research on weakly supervised object detection algorithms,most methods transform the object detection problem into the classification problem of region proposals,which makes the weakly supervised object detector tend to locate the salient and discriminative object part areas.This paper proposes a weakly supervised object detection method that combines attention and erasure mechanism.By using the attention map to search for the region with greater discrimination in the region proposal,and then using the erasure mechanism to erase the region,forcing the detection model to enhance the learning of weaker discriminative region feature.Secondly,the weakly supervised object detection algorithm based on multi-instance learning selects the region proposal with the highest class score as pseudo object,the scenarios of multiple objects of the same category.To solve the problem,we proposed an improved pseudo object estimation method.In the early stage of training,we only select the region proposal with the highest class score as the pseudo object.With the increase of model iterations,the class confidence of object gradually increases,when the class confidence greater than the threshold,transfer to multiple object scene.Finally,in order to improve the localization ability of the detector,this paper cascades the weakly supervised object detection network and the fully supervised object detection network,designs an end-to-end object detection model,and uses the prediction of the weakly supervised object detector as pseudo label to train the fully supervised object detector online.We train the weakly supervised object detector and fully supervised object detector in a multi-task learning manner.In this paper,the comparison and verification are carried out on two general data sets.The experimental results show that the method proposed in this paper can effectively improve the detection accuracy of the weakly supervised object detector.Compared with the baseline methods,the mean average precision of each class and the localization ability of the detector have been greatly improved.