A Deep Learning Method for Detecting Floating Garbage in Urban Rivers

In recent years, plastic pollution in the ocean has become a global environmental problem, deeply affecting the ecosystem as well. In fact, 80 percent of ocean plastic was reported to come from terrestrial river basins, therefore it would be extremely important to recognize how much plastic waste runoff from rivers around the world was. In response to this global environmental problem, an image analysis method for monitoring river waste transport has recently been started to propose. However, this image analysis indicated a difficulty to fully detect plastic waste in various types of waters because it needed to use a color difference in each water to classify the type of waste.

This paper tried to develop an automated detection system based on a modified convolutional neural network (CNN) for detecting and counting floating river waste. The CNN used in this research was You Only Look Once (YOLO) architecture with a fine-tuning for adjusting it to the waste detection. The proposed model has further been improved its accuracy through the enlarged image processing. As for the waste counting, an object tracking method, e.g., deep SORT, could be used with the proposed model in video frames of flowing water.  

The results showed that the proposed YOLO model with enlarged image processing achieved the evaluation values of mean average precision mAP (%) of can, carton, plastic bottle, foam, glass, paper, and plastic were 95, 89, 94, 97, 92, 71, and 81, respectively. Moreover, the proposed mode with deep SORT has achieved the F1-score (%) of 80, 80, 75, 85, 100, 100, and 50, in each waste type. Consequently, the proposed model could be feasible for identifying and counting flowing river waste accurately. The future research work should improve the counting accuracy and further develop an automated model implementation method.