Enhanced Small Target Detection in Optical Remote Sensing Images using CAC-RetinaNet with Multi-Scale Feature Fusion and Context Attention

Small target detection in optical remote sensing images poses a significant challenge due to the complex background and scarcity of small target features. This paper proposes a novel CAC-RetinaNet model to address this issue.The model comprises three key components: 1) A new CCB-Net backbone network incorporating the Criss Cross-Attention module to enhance context information extraction and long-range dependency modeling.2) An HSA-FPN feature fusion module utilizing ASPP to enlarge the receptive field of convolutional kernels, thereby improving the detection of small targets. 3) A cascaded sparse convolution query (CSQ) mechanism in the detection head to reduce computational overhead associated with high-resolution layers. Extensive experiments on DIOR and LEVIR-SHIP datasets show that our method achieves state-of-the-art performance, with mAP achieving 2.25 % and 0.53 % improvements on the dataset, respectively, compared to the latest SIFANet method. These results validate the effectiveness of the proposed CAC-RetinaNet model for small target detection in optical remote sensing images.