Underwater Small Target Detection Algorithm Based on Deep Learning

Underwater small target detection has broad application prospects in underwater resource exploration, underwater engineering operations, and underwater biodiversity assessment. A new underwater target detection method combining a parameter-free attention mechanism and a hybrid model is proposed to address issues with existing models in underwater small target detection. This method first integrates the E-ELAN module and the parameter-free attention mechanism in the feature extraction part of the backbone network; then, in the inference prediction stage of the head network, the ENAN-H module containing the hybrid model is introduced to replace the traditional standard convolution layer. Additionally, the SIoU loss function is used for boundary box loss calculation, further improving detection accuracy and reducing error rates. A series of experiments conducted on the augmented URPC dataset demonstrate that the proposed detection method increases the mean accuracy by 23.1 % compared to the YOLOv7 model, showing superior performance with fully validated effectiveness and feasibility.