基于形状关系的目标检测模型异步检测头构建方法

Construction method of asynchronous detection head for object detection model based on shape relation

从目标检测网络输出的特征图与输入目标之间具有的形状关系出发,提出一种目标检测模型异步检测头构建方法,先进行分类预测,再添加边框预测网络使之利用形状关系预测出目标框的检测头结构。这一结构将分类网络与边框回归网络完全分开,避免了特征不匹配问题。在添加为保留形状关系而改进的LOSS函数后,使用该检测头结构的MobileNet-YOLO目标检测网络在变电设备典型缺陷数据集上不同类别的精确率提升了2%~14%。在对结果的分析中推测检测头的适应场景,并使用VOC数据集对推测进行了验证。最后对检测头的有效性进行了分析,并指出了算法适应的数据集类型。

Starting from the shape relationship between the output characteristic graph and the input object,a method of constructing asynchronous detection head is proposed in this paper for object detection model,which first performs classification prediction,and then adds border prediction network to predict the structure of the detection head using the shape relationship.This structure completely separated the classification network from the border regression network and avoids the feature mismatch problem completely.After adding the improved loss function to preserve the shape relationship,the Mobilenet-YOLO object detection network based on this detector structure improved the accuracy by 2%-14%in different categories of typical defect data sets of substation equipment.In the analysis of the results,the adaptive scene of the detection head was speculated,and the prediction was further verified by VOC data set.Finally,the validity of the detection head was analyzed,and the type of data set that the algorithm adapts to was pointed out.