High precision object detection in remote sensing images by combining feature enhancement and anchor automatic generation

Object detection is an very important and challenging tasks in the field of remote sensing image processing. To solve the problem of low accuracy of multi-scale object detection in remote sensing image caused by differences of object scale and random directional distribution, this paper proposed a object detection method based on feature enhancement and anchor automatic generation module. In this method, a controllable atrous convolution module is added into the Resnet50 network, and based on this, an enhanced feature pyramid network is designed to improve the multi-scale feature representation ability of the network. In the region generation network, the anchor automatic generation module is used to learn the position and shape of the anchors independently, so as to obtain more sparse and high quality candidate regions. The comparison experiments and analyzes between proposed method and some other detection algorithm based on convolution neural network on the NWPU VHR-10 dataset and aircraft object dataset, the results show that the mAP of the proposed method is optimal on the two datasets, which are 99.2% and 87.7%, respectively. This method has strong scale adaptive ability and effectively improves the accuracy of multi-scale object detection in remote sensing images.