Remote Sensing Object Detection in Disturbed Environment Based on Improved RTMDet Network

Remote sensing object detection remains a challenging task in computer vision due to the arbitrary angles of the targets and the complex environmental conditions, which contribute to low detection accuracy. While numerous scholars have concentrated on aligning horizontal bounding boxes with rotation bounding boxes, altering the generation method of rotation bounding boxes, or improving the loss function, they have overlooked the utilization of prior knowledge in remote sensing target detection. In our paper, we introduce a new network structure, Large Kernel Cross-Stage-Partial-connection Layer (LK-CSPLayer), based on the CSPLayer module of Real-Time Models for Detection (RTMDet), leveraging prior knowledge in remote sensing target detection. Given the presence of remote sensing targets at relatively long distances, the recognition of these targets, especially smaller ones, relies on the surrounding background information. We enhance the accuracy of network detection by expanding the convolutional kernel to augment the receptive field and extract extensive contextual information. Furthermore, during the operation of remote sensing satellites, environmental factors can lead to unclear remote sensing images, consequently impacting the quality of feature extraction. To address this, we propose a series of image preprocessing methods. To verify the effectiveness of our approach, we construct a disturbed remote sensing dataset and assess the method’s efficacy. Our experiments on the public dataset high resolution ship collections 2016 (HRSC2016) demonstrate that utilizing LK-CSPLayer can achieve a 2.0% improvement on the baseline, while experiments on our custom disturbed dataset reveal that utilizing our preprocessing methods and LK-CSPLayer both lead to a 1.6% improvement on the baseline.