Find in Library

Due to the long distance and large-scale of satellite imaging, and the high complexity of depth convolutional neural network, common detectors cannot be directly applied to large-scale remote sensing images. Therefore, this article proposes a two-phase object detection network from key points to region proposals, namely P2RNet. In the first phase, the key points of all suspected objects are obtained through the key point extraction network, and then these key points are divided into multiple region proposals using the region proposal generator. In the second phase, these region proposals are input into the lightweight object detection network to achieve fast and accurate maritime object detection. The lightweight object detection network is improved based on YOLOv5. To significantly reduce the number of parameters and computation of the network, the improved MobileNetv2 constructed by the grouped sandglass block is used as the backbone to extract sufficient feature information. To effectively improve the detection accuracy of the network, the simple attention module is embedded in the feature fusion network to strengthen the feature fusion process, and the oriented spatial pyramid pooling-fast is proposed to capture long-distance dependencies. The experimental results on the DOTA Ship dataset show that the average precision and frames per second of the lightweight object detection network reached 80.59% and 112, respectively, achieving a good balance. Moreover, the overall object detection network achieved excellent detection results on large-scale remote sensing images.