Surface Water Garbage Detection, Classification and Counting Using Modern Image Recognition
Abstract
Prathamesh Garud
Garbage floating on rivers, lakes and ponds harms water life, blocks drains and spoils the look of public places. Checking it by eye is slow and tiring. This paper builds a simple image-based system that finds floating garbage in a photo, tells what type it is, and counts how many pieces of each type are present. A small YOLOv8 model (YOLOv8n, about 3.0 million parameters) was trained on a public water-garbage image set with 371 training images, 105 validation images and 7 labels. Counting is done by adding up the boxes the model draws for each class. In the training log used for this paper (24 completed epochs), the model reached a best mAP50 of 0.244, with precision of 0.624 and recall of 0.276 at that point, and it took about 155 ms to process one image on a CPU. These scores are modest. They show that the idea works even with very little data and no GPU, and they point clearly to what must improve: more labelled images, cleaner class labels, and a longer training run.

