Enhanced Remote Sensing Object Detection Algorithm Based on YOLOv7

Remote sensing object detection faces challenges such as densely distributed objects, complex backgrounds, and significant variations in object scale, which result in issues like missed detections and false positives. Additionally, the detection process suffers from long-range dependencies, leading to the loss of detailed spatial semantic information. In this study, we present an enhanced YOLOv7-based algorithm for remote sensing object detection. Initially, deformable convolution layers, specifically Dynamic Convolution (ODConv), are incorporated to bolster the network's feature extraction capabilities. Subsequently, the ELAN module is improved through the implementation of the SimAM attention mechanism, mitigating to some extent the issues of long-range dependencies and the resultant false positives and negatives caused by complex backgrounds and dense object distributions. Finally, the original loss function is optimized using an EIOU loss function, thereby improving the network's localization accuracy and accelerating model convergence. On the DOTA dataset, the modified algorithm achieves a mean average precision (mAP) of 0.5 at 67.5%, representing a 4.9% improvement compared to the original YOLOv7 model.