A Semi-Supervised Object Detection Algorithm Based On Teacher-Student Models With Strong-Weak Branches

Object detection technology is widely used and performs well in many fields such as autonomous driving,security surveillance,and image search engines.Highperformance object detectors benefit from large-scale high-quality annotated datasets.The construction of high-quality object detection datasets relies on manual labeling,which is inefficient and resource-intensive.Semi-supervised object detection algorithms can train object detectors using a small fraction of labeled data and a large amount of unlabeled data.This approach significantly reduces the manual labeling workload and improves the performance and generalization ability of the model to a certain extent.Based on the research value of the semi-supervised object detection algorithm,this paper proposes a semi-supervised object detection algorithm based on the teacherstudent model with strong and weak branches,which works as follows.(1)Applying the teacher-student model framework to the semi-supervised object detection problem.The teacher model of the conventional semi-supervised object detection algorithm generates all pseudo labels at once,and this approach leads to no way to update the pseudo labels during the training process,which in turn limits the overall algorithm performance.In the teacher-student model framework of this paper,the exponential sliding average is used to update the parameter of the teacher model,which further removes the performance limitation of the fixed pseudo labels on the model.(2)Improving the loss function to solve the category imbalance problem.In semisupervised object detection,category imbalance is manifested as foregroundforeground imbalance,and this sample difference can significantly affect the final object detection performance.In this paper,Focalloss loss is used to alleviate this imbalance problem.(3)A two-branch structure of the teacher model is proposed to solve the pseudo label localization quality measurement problem.In the semi-supervised object detection problem,the confidence level comes from the output of the classifier,which can be used to filter the classification labels but should not be used to filter the localization labels.In this paper,the quality measurement problem of pseudo label localization is solved by the strong and weak double branch structure of the teacher model to obtain higher quality pseudo labels.(4)The student model double branch structure is proposed to solve the problem of reasonable training of pseudo-labels.Mixing noisy pseudo labels and ground truth labels in one branch for training may cancel out the correct parameters learned from the ground truth data,which leads to performance degradation.In this paper,the strong and weak branches of the student model are decoupled to reduce the negative impact of noise in pseudo labels on classification and regression.This paper proposes a semi-supervised object detection algorithm based on the FasterRCNN framework,i.e.,a semi-supervised object detection algorithm based on the teacher-student model with strong and weak branches,which has superior performance.The method in this paper achieves 52.5 mAP(+1.8)on the PASCAL VOC dataset and 53.5 mAP(+3.2)using MS-COCO train 2017 as additional unlabeled data.On the MS-COCO dataset,the method in this paper also improves about 1.0 mAP with 10%of COCO and COCO-full as the experimental configuration of labeled data.