AK-YOLOv9: Improved Object Detection for Pollutants in Rivers Discharging into the Sea

Abstract Pollutants transported by rivers into the marine environment significantly affect the ecological health of coastal waters. In the context of the Pinglu Canal construction, we developed an enhanced target detection model, AKS-YOLOv9, based on YOLOv9 for monitoring pollutants in riverine discharges to the sea. This model improves the network's feature extraction capabilities for irregular and multi-scale targets. The key enhancements include: (1) Incorporating the AKConv module into the detection head of YOLOv9 to enable feature extraction using irregular convolution kernels; (2) Integrating the lightweight self-attention mechanism SimAM within the backbone of YOLOv9, thereby improving the network's perceptive ability for spatially distributed features and contextual relationships; (3) Employing the MPDIOU function to compute the bounding box regression loss, which serves to boost the efficiency and accuracy of bounding box regression. Ablation study results indicate that compared to the original model, the improved model achieves a 2.7% increase in accuracy and a 2.1% improvement in recall rate, further validating its reliability and effectiveness. These findings hold significant promise for environmental monitoring applications in riverine systems discharging into the sea.