降低冗余检测框数量的目标检测算法

Object detection algorithm for reducing the number of redundant detection bounding boxes

在机器视觉领域中,很多无锚框检测算法在处理目标密集的图像时会产生冗余边界框的现象,降低了检测精度。针对这种现象,本文借助RetinaNet的网络结构提出一种可以降低冗余检测框数量的目标检测方法。首先在特征提取阶段,加入一种新的注意力机制来提高特征的表达能力;然后为了减少正样本中标签错误标定的可能,对选取正样本的位置进行筛选,之后将算法选择的正样本输入预测分支得到目标边界框的坐标和置信度;最后根据目标边界框的位置和分类结果,提出一种类内的交并比分数重分配推理策略,该策略能够减少重叠的检测框数量,从而提高算法精度。本文算法的有效性在公开的图像数据集上进行验证,结果表明,所提出的算法可以提高检测精度、优化定位效果,具有较好的应用前景。

In the field of machine vision,many anchor-free object detection algorithms produce redundant bounding boxes when processing dense images,which reduces the detection precision.In view of this phenomenon,a target detection method is proposed which can reduce the number of redundant detection frames by using the network structure of RetinaNet.Firstly,in the feature extraction stage,a new attention mechanism is added to improve the expression ability of features.Then,in order to reduce the possibility of wrong label calibration in the positive samples,the position of the selected positive samples is filtered,and then the positive samples selected by the algorithm are input into the prediction branch to obtain the coordinates and confidence of the object bounding box.Finally,according to the location and classification results of the object bounding boxes,a new strategy of intersection and fraction redistribution is proposed,which can reduce the number of overlapping detection boxes and improve the precision of the algorithm.The effectiveness of this algorithm is verified on the open image data set.The results show that the proposed algorithm can improve the detection precision and optimize the positioning effect,and has a good application prospect.