Research on an SSD remote sensing image object detection algorithm based on HSIAM

Aiming at the problems, such as missed detection of small targets, positioning deviation of rotating targets, and complex background interference in remote sensing images, an improved SSD algorithm based on the High-Level Semantic Information Activation Module (HSIAM) and the improved BSWIoU based on Bhattacharyya distance was proposed. The HSIAM module enhances information fusion capabilities within the deep network. The CA mechanism employs adaptive average pooling to enhance focus on central regions of feature maps, distinguishing small targets within complex backgrounds. The RBD_IoU loss function integrates an orientation-matching constraint and a dynamic weighting mechanism to mitigate rotational bounding box regression bias. Experimental results for three benchmark datasets (DIOR, DOTA, and NWPUCHR) showed that, compared with the baseline SSD algorithm, the mAP50 of the improved model increased by approximately 2%. Furthermore, it achieved a balanced trade-off between accuracy and speed, with 12.5% fewer parameters than YOLOv8s. This provides a high-precision and lightweight solution for target detection in remote sensing images.