An efficient network for object detection in scale-imbalanced remote sensing images

Abstract Remote sensing images contain complex scene information and have multi-scale characteristics after being imaged by remote sensing equipment because the target instances in natural scenes are different sizes. In addition, small target instances are more difficult to identify in complex backgrounds, resulting in serious scale imbalance problems, which pose a serious challenge to identifying and positioning target objects. To address this problem, this article proposes a detection method for multi-scale remote sensing image targets by combining the frequency attention mechanism. The model first introduces an improved frequency channel attention mechanism to design a feature extraction module to improve the extraction of key features by the neural network; second, considering that the complete intersection over union method does not comprehensively consider the aspect ratio of the bounding box, which will cause the loss of small-scale target feature information, the efficient intersection over union method is used to improve it; then, because of the high missed detection rate of the non-maximum suppression (NMS) method, Soft-EIoU-NMS is used to replace the original NMS method. The experiment first conducted ablation experiments on the LEVIR dataset, where the target scale changes little and the number of ground object categories is small, and the DIOR dataset, where the target scale changes greatly and the number of ground object categories is large. The [email protected] on the LEVIR dataset reached 0.935; the [email protected] on the DIOR dataset reached 0.882. Then, the model proposed in this article was compared with the mainstream target detection methods. The experimental results verified the effectiveness of the model in this article. Finally, the model was applied to the disaster remote sensing image scene for detection, again demonstrating the model’s good detection performance. Therefore, the experimental results show that the method proposed in this article can effectively alleviate the scale imbalance problem and achieve a good target detection effect.