Multi-scale Context-Aware Detection model for Underwater Target

To address the issues that traditional models cannot effectively handle the complex underwater environmental noise, the target scale varies greatly, and the inability to balance model size and accuracy, MSCA-UODA(Multi-scale Context-Aware Underwater Object Detection Algorithm) model was proposed. The model designs a context-enhanced downsampling module, CEADown (Context Enhanced ADown), which can effectively reduce the model parameters, capture context information efficiently and reduce underwater environmental noise. This model also proposes a multi-scale feature extraction module based on dual-path partial connection, named CSP-MSPF(Cross Stage Partial-multi-scale Partial Feature), and uses SHSA (Single-Head Self-Attention) mechanism to improve C2PSA, enhancing the multi-scale feature extraction capability of the model. Through experiments, MSCA-UODA improved by 2.0% and 1.1% respectively on mAP50 compared with baseline model on the datasets URPC2020 and DUO. The number of parameters decreased by 12.01%, and its comprehensive performance was superior to the current mainstream object detection models.