A Hybrid VGG-ResNet Feature Fusion Network for Object Detection in Side-Scan Sonar Images

Side-scan sonar images present significant challenges for object detection due to high noise levels, limited spatial resolution, and complex seabed structures. These factors, along with speckle noise and acoustic shadowing effects, further complicate reliable target detection. Nonetheless, object detection is extremely important in many areas, including marine archaeology, underwater search and rescue, mine countermeasures operations, and the inspection of critical national infrastructure, to improve safety and operational efficiency. In this work, a hybrid convolutional neural network (CNN) is presented for object detection in challenging side-scan sonar imagery. The proposed model combines the complementary feature-extraction capabilities of VGG-16 and ResNet-50 via feature fusion to improve target discrimination in sonar images. Transfer learning from ImageNet-pretrained backbones is employed to address data sparsity and improve model generalization on sonar datasets. Experimental evaluation demonstrates that the proposed model achieves an overall classification accuracy of 84.2% and a mean Average Precision (mAP) of 88.35%, outperforming several existing methods. The results pave the way to enhance the efficacy and accuracy of underwater surveys, search-and-rescue missions, and seabed mapping.