A water surface garbage recognition method based on transfer learning and image enhancement

Abstract

In this paper, we propose a solution to the escalating issue of water pollution caused by garbage in rivers, reservoirs, and oceans. We introduce the implementation of the You Only Look Once version 5 (YOLOv5) algorithm to address the challenges associated with identifying garbage targets, such as light reflection and varying shooting angles. Our algorithm effectively tackles the issue of identifying small and variable-sized targets, while incorporating image enhancement techniques to enhance the adaptability of the target recognition algorithm to different lighting conditions for accurate identification of water surface garbage. To prevent overfitting, we employ the k-fold cross-validation method, and further enhance the algorithm's performance through transfer learning, ensemble learning, and test-time augmentation (TTA). During the ensemble learning stage, we evaluate a half-precision model to strike a balance between prediction speed and accuracy, ultimately achieving an optimal solution. Using our self-made dataset, our study demonstrates a detection speed of 67.7 ms and a mean average precision at 50% IOU (mAP50) of 98.1%. These results exhibit practical application value, surpassing other deep learning methods in this field.

Keywords

• Water surface garbage recognition
• High dynamic range (HDR)
• Histogram equalization
• Transfer learning
• Ensemble learning
• K-fold cross-validation