Reserch On Semi-supervised Learning For Object Detection

In recent years,with the continuous development of deep learning,the performance of various tasks in the field of computer vision has skyrocketed,but these developments rely to some extent on large-scale,labeled data sets.However,data labeling is a costly task.Compared with image data acquisition,it is cheaper to acquire only image data,and there are many ways to acquire it and a wide range of sources.Therefore,how to utilize unlabeled image data is a hot research issue recently.Thesis studies how semi-supervised object detection algorithm uses a small amount of labeled data and a large amount of unlabeled data to achieve supervised learning effect.The main work is as follows:Aiming at the problem that the existing single-branch teacher-student structure cannot fully mine the unlabeled data information,thesis proposes cross-supervised co-training of the two-branch teacher-student structure.The differences between the two models can provide more useful information for each other for network learning,achieve effective decoupling,and fully mine the information contained in the unlabeled data.In addition,Focal Loss was used to replace cross-entropy loss function and combined with MUM data enhancement strategy to alleviate the prediction bias caused by class imbalance in target detection data set.Experimental results show that compared with the traditional single-branch structure algorithm,the accuracy of the two-branch structure is improved by 7%,which verifies the effectiveness of the two-branch teacher-student structure.Aiming at the quality problem of pseudo-labels generated during model training,based on the proposed two-branch teacher-student structure,a pseudo-label generation mechanism is further proposed.Thesis extracts a more refined feature representation by combining the same-level feature fusion,and matches the output of the double-branch according to certain rules to get a bounding box matching pair.If the matching decisions are consistent,the output is a pseudo-label.If they are inconsistent,the system filters the threshold.If the threshold is higher,false labels are displayed.The experimental results show that compared with the existing two-branch structure that only uses threshold filtering,the pseudo-label generation mechanism improves the performance by 1.56%.Finally,the proposed algorithm is applied to the automatic driving scene and mobile detection terminal,and good performance is obtained,which shows the effectiveness of the proposed semi-supervised target detection algorithm.