Water Surface Highlight Compensation Method Using Feature Fusion and Improved Criminisi Algorithm

In surface-floating debris detection tasks, intense highlights on the water surface can degrade images and occlude floating objects, leading to missed and erroneous detections that compromise the accuracy of detection algorithms. This study proposes a highlight detection and compensation method utilizing feature fusion and an improved Criminisi algorithm. First, the mechanism of water surface highlights is analyzed, and a water surface highlight dataset is collected. The color and positional information of highlighted pixels are combined to train a Fully Convolutional Network (FCN) model for identifying highlight areas, excluding highlights in the sky. For the detected highlight regions on the water surface, an improved Criminisi algorithm is applied to compensate and restore the water surface images. Finally, the accuracy of floating debris detection is validated using the YOLOv8 algorithm. Results show that the mean Average Precision (mAP) after highlight compensation reaches 88.7%, representing an increase of 1.7% compared to images without highlight compensation, with [email protected]:0.95 improved by 2.2%. In water surface multi-object detection, significant improvements have also been achieved.