ROI-YOLO Model: Real-Time Deep Learning Detector for Floating Objects on River

Floating objects on river and lake surfaces can cause significant water pollution, necessitating timely detection and treatment. However, the features of floating objects are complex and variable, with a small proportion in monitoring scenes and image quality constrained by surveillance equipment. To address these challenges, this study proposes the ROI-YOLO model for rapid and intelligent detection of floating objects in monitoring video images. The proposed model comprises three key components. First, the pre-detection module enables fast scene focusing and background suppression for large-scale video images, overcoming the issue of excessive information loss caused by downsampling in real-world scenarios. Second, the model employs multi-resolution feature extraction and precise localization of floating objects, with the design of adaptive anchor boxes based on object density and maximum-minimum distances. Third, the sub-pixel feature fusion module learns upsampling operators on feature maps and integrates multi-resolution features based on sub-pixel information, enhancing the restoration and fusion of effective information to improve detection accuracy. Furthermore, the model utilizes Bernoulli distribution-based random parameter updates during training, reducing the required training sample size. The resulting model achieves a mean average precision of 84.5%, a recall of 90.9%, and a detection speed of 27 FPS (Frames Per Second), outperforming other deep learning based object detection models.