【摘要】 传统的半监督目标检测方法在使用固定阈值时存在着一些局限性，如偏差和标签分布不均衡等问题。针对标签分布不平衡问题，提出了一种局部特征增强的自适应方法来捕捉图片的细节特征。首先，在通道注意力模块中加入特征提取模块，对输入特征进行二次特征提取，得到更多的细节特征；其次，引用了自适应标签分布感知置信阈值方法，根据所设置的置信度阈值，自适应地对标签分布进行调整，解决标签预测任务中的标签分布不均衡问题。在COCO(Common Objects in Context)数据集上的实验结果表明，本算法可以解决模型对少数类别图片的预测能力不足的问题，并达到较好的检测性能。更多还原

【Abstract】 Traditional semi-supervised target detection methods have some limitations when using fixed thresholds, such as bias and unbalanced label distribution. For the problem of label distribution imbalance, an adaptive method of local feature enhancement is proposed to capture the detailed features of pictures. First, a feature extraction module is added to the channel attention module to perform secondary feature extraction of the input features to obtain more detailed features; second, the perceptual confidence threshold method of adaptive label distribution is cited, and the label distribution is also adaptively adjusted according to the set confidence threshold to solve the problem of unbalanced label distribution in the label prediction task. Experimental results on the COCO dataset show this method can solve the problem of insufficient model predictive power for a few categories of pictures and achieve good detection performance.更多还原