Object Detection in Remote Sensing Imagery Based on Prototype Learning Network With Proposal Relation

Abstract

Deep learning object detection algorithms, due to their powerful feature learning capabilities, can effectively improve the accuracy of target detection in remote sensing images. However, remote sensing image target detection faces challenges such as dense arrangement of small targets and complex backgrounds. Addressing the above issues, how to enhance the receptive field while effectively depicting the structural relationships between proposals will be beneficial for detecting small targets in remote sensing images with complex backgrounds. Motivated by this, a prototype learning network with proposal relation, called PLNet-PR, is proposed for remote sensing object detection, while enhancing receptive fields. The shift operation is inserted into the inception module and spatial graph convolution layer, constructing sparse shift selective convolution (S3Conv) based on spatial-channel selective attention mechanism, and graph-guided proposal-relation learning module (GPRLM), for enhancing the characterization of small targets and acquiring powerful proposal-level feature relations of remote sensing targets. Furthermore, a category prototype repository (CPRep) with a class-wise semantic attention (CWSA) block is proposed for the improved proposal generation between different remote sensing object categories. Our extensive experiments validate the effectiveness of PLNet-PR which outperforms other related deep learning methods. Codes are available: https://github.com/RSIP-NJUPT/PLNet-PR .