The variable anchor box region proposal network based on confidence non-maximum suppression

摘要

The traditional target detection algorithm based on manual features has achieved good results in the past time, but the computation amount and running speed of the algorithm are not satisfactory. At present, the algorithm based on deep convolutional neural network has become the first choice and widely used. In this paper, attention guidance module is introduced in the detection based on the deep convolutional neural network, which guides the anchor box of region proposal network (RPN), making the selection of anchor box shape and size more effective. At the same time, a confidence factor non-maximum suppression (NMS) method is proposed to solve the problem of false detection and missed detection in the traditional post-processing, which makes a great contribution to the overall performance of the model. In the experiment, we found that our method has good detection performance in both RPN variants and existing advanced algorithms.