Improved YOLOv7 Underwater Object Detection Based on Attention Mechanism

The task of detecting marine target organisms has always been a challenging issue, despite the numerous machine learning detection methods proposed to improve precision. The underwater image blurriness caused by irregular light absorption and water quality remains a major obstacle to achieving accurate detection. This results in high misalignment rates and poor underwater scene recognition capabilities for detecting underwater targets. To address this, we put forward a YOLOv7-RNCA underwater target detection technology based on improvements to YOLOv7. This model adds residual modules and coordinate attention mechanisms (CA) at the end of the backbone network, as well as incorporating partial convolution (PConv) modules. The combination of these three components makes the model more precise during the detection process while reducing unnecessary computation and memory access. This allows for better optimization during deep network training and preserves more feature information. Additionally, we reconstructed the SPPCSPC structure and incorporated a global attention mechanism (GAM) to form the SPPCSPC-GAM module in the neck network, which improves the performance of the convolutional neural network (CNN) and ensures good data capabilities and robustness during training, thereby enhancing the target detection ability. We also improved the neck ELAN module by introducing PConv convolution modules, which continuously enhance network learning abilities without disrupting the original gradient path. The introduction of the PConv module reduces redundant computation and memory access, making the ELAN-PConv module more effective at extracting spatial features. Our outcomes of experimentation indicate YOLOv7-RNCA network an average precision of 86.6% on the URPC dataset, outperforming existing methods in accuracy detection and demonstrating great potential as a promising solution for marine target monitoring tasks.