A One-Stage HMDV Algorithm Applied in Multitarget Detection in SAR Images

Synthetic aperture radar (SAR) image target detection is a key method in image interpretation. Currently, most radar target detection methods are primarily designed for single category, without adequately addressing the challenges of multitarget detection accuracy and lightweight deployment in coastal areas. To tackle these issues, this article proposes a single-stage multi-target detection network called HMDV. First, a hybrid feature extraction module is designed to address the computational complexity caused by increased width and depth in convolutional neural networks. Second, to overcome the challenges posed by the diversity of target sizes in multi-target detection, the loss of dense target feature information, and background clutter, a mult-idimensional perceptual feature aggregation and dispersion module is developed. This module effectively improves the detection accuracy for aircraft and oil tank targets in SAR images. Finally, to resolve the issue of low detection performance due to the small proportion of small targets in the prediction boxes, a new width and height vector loss function is proposed. This function simultaneously constrains the width, height, and proportion of bounding boxes, enhancing the network’s convergence speed and reducing misdetections of small-sized ships. Experimental results demonstrate that the proposed model improves mean average precision accuracy by 2.4% and reduces the number of parameters to 13% of the original, confirming the model’s effectiveness and robustness.