Multiscale Object Detection in Remote Sensing Images Combined with Multi-Receptive-Field Features and Relation-Connected Attention

Object detection is an important task of remote sensing applications. In recent years, with the development of deep convolutional neural networks, object detection in remote sensing images has made great improvements. However, the large variation of object scales and complex scenarios will seriously affect the performance of the detectors. To solve these problems, a novel object detection algorithm based on multi-receptive-field features and relation-connected attention is proposed for remote sensing images to achieve more accurate detection results. Specifically, we propose a multi-receptive-field feature extraction module with dilated convolution to aggregate the context information of different receptive fields. This achieves a strong capability of feature representation, which can effectively adapt to the scale changes of objects, either due to various object scales or different resolutions. Then, a relation-connected attention module based on relation modeling is constructed to automatically select and refine the features, which combines global and local attention to make the features more discriminative and can effectively improve the robustness of the detector. We designed these two modules as plug-and-play blocks and integrated them into the framework of Faster R-CNN to verify our method. The experimental results on NWPU VHR-10 and HRSC2016 datasets demonstrate that these two modules can effectively improve the performance of basic deep CNNs, and the proposed method can achieve better results of multiscale object detection in complex backgrounds.

Automated summary

This article proposes an object detection algorithm for remote sensing images based on multi-receptive-field features and relation-connected attention. The algorithm utilizes dilated convolution to aggregate context information from different receptive fields, allowing for better adaptation to object scale changes in complex scenarios. The inclusion of a relation-connected attention module, which combines both global and local attention, enhances feature discriminability and detector robustness. Experimental results demonstrate that these modules effectively improve the performance of basic deep CNNs and achieve better results in multi-scale object detection in complex backgrounds.