Abstract: To address the challenges posed by large-scale small object detection, dense object distribution, and issues of missed detections and false positives in remote sensing images, this paper introduces a remote sensing image object detection approach grounded in an enhanced YOLOv7 model. Initially, the method integrates the DCNv2 structure and residual architecture into the YOLOv7 model, reconstructing a novel backbone network to enhance the extraction of shallow-level features information and improve network accuracy. Subsequently, a pioneering feature fusion module is incorporated into the neck network, combined with the SimAM mechanism, which adaptively adjusts the fusion weights of both shallow-level texture information and deep-level semantic information, thereby effectively curbing noise introduced during shallow feature extraction and augmenting the representation of essential features. Finally, the normalized Gaussian Wasserstein distance loss function is used as the regression loss function, replacing the traditional IOU to improve the detection capability for multi-scale targets. Empirical findings derived from the DOTAv1.0 dataset reveal an average precision of 20.1% for small objects, while the DIOR dataset yields an average precision of 29.0%. Furthermore, compared to recent advanced methods such as YOLOv7 and YOLOv6, the proposed algorithm demonstrates strong competitive performance.

Key words: remote sensing image, object detection, deformable convolutional networks, SimAM attention mechanism, Gaussian Wasserstein distance