关键词: 半监督目标检测, 自步学习, 一致性正则化, 动态自步参数, 连续权重变量

Abstract: In order to improve the quality of pseudo-labels and solve the problem of confirmation bias in semi-supervised object detection,a semi-supervised object detection framework based on dynamic parameters under the guidance of self-step learning was proposed.The framework was designed with dynamic self-stepping parameters and continuous weight variables to optimize the effect of semi-supervised object detection.The dynamic self-stepping parameter evaluates the difficulty of the sample based on the real-time performance of the model during the training process.The continuous weight variables accurately evaluated the importance and reliability of each sample in training by comparing the relationship between sample loss and dynamic self-stepping parameters,and refined weight design was carried out for each object in the sample.In addition, the framework uses a single model for iterative training, and introduces a consistency regularization strategy to evaluate the consistency of model predictions.This design not only provides more targeted weight information for the model, but also adaptively optimizes the training process through the dynamic adjustment of the weight information.Extensive comparative experiments on PASCAL VOC and MS-COCO datasets show that the framework not only significantly improves the detection accuracy of the model, but also verifies its wide versatility and efficient convergence performance.

Key words: Semi-Supervised Object Detection, Self-Paced Learning, Consistency Regularization, Dynamic Self-Paced Parameter, Continuous Weight Variable