Due to the large number of parameters and high computational complexity of current target detection models, it is challenging to perform fast and accurate target detection in side-scan sonar images under the existing technical conditions, especially in environments with limited computational resources. Moreover, since the original waterfall map of side-scan sonar only consists of echo intensity information, which is usually of a large size, it is difficult to fuse it with other multi-source information, which limits the detection accuracy of models. To address these issues, we designed DBnet, a lightweight target detector featuring two lightweight backbone networks (PP-LCNet and GhostNet) and a streamlined neck structure for feature extraction and fusion. To solve the problem of unbalanced aspect ratios in sonar data waterfall maps, DBnet employs the SAHI algorithm with sliding-window slicing inference to improve small-target detection accuracy. Compared with the baseline model, DBnet has 33% fewer parameters and 31% fewer GFLOPs while maintaining accuracy. Tests performed on two datasets (SSUTD and SCTD) showed that the mAP values improved by 2.3% and 6.6%.
Read full abstract