State-of-the-Art Image Recognition for Surface Water Garbage Detection, Classification, and Counting
This paper presents a lightweight YOLOv8n-based image recognition system for detecting, classifying, and counting surface water garbage, demonstrating that while the approach is feasible with limited data and CPU resources, its current modest performance (mAP50 of 0.244) highlights the need for expanded datasets, refined labels, and extended training to achieve practical utility.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Rivers, lakes, and ponds are the veins of our landscape, but they are increasingly clogged with a different kind of debris: floating trash. Plastic bottles, bags, and other waste drift on the surface, trapping wildlife, blocking drainage, and ruining the view. For now, the only way to track this pollution is for people to walk the banks or ride boats, squinting at the water and writing down what they see. This method is slow, inconsistent, and impossible to repeat often enough to catch the problem as it grows. The solution lies in giving machines the ability to see. By mounting cameras on bridges, boats, or drones, we could capture images of the water every day. The challenge is teaching a computer to look at a single photograph, find every piece of garbage, identify what it is, and count how many pieces are there. This is the work of object detection, a field where artificial intelligence learns to recognize shapes and patterns just as a human eye does, but with the speed and endurance of a machine.
A researcher at Sanjivani University in India set out to test whether a small, efficient computer program could handle this task without needing expensive, high-powered hardware. The goal was to build a system that could find floating trash in a photo, name the type of waste, and tally the total count, all while running on a standard laptop processor. The researcher chose a specific tool known as YOLOv8, a modern detection system designed to be fast and lightweight. To teach this system, they used a collection of 371 images of water surfaces containing garbage, along with 105 additional images to check the system's progress. These images were marked with labels for seven different categories, ranging from plastic bottles and cans to fiber waste and other debris, though some of the labels in the original data set were imperfect placeholders left over from the labeling process.
The training process involved feeding these images to the computer model over a series of steps called epochs, allowing the software to adjust its internal rules to better recognize the trash. The entire training run took place on a central processing unit, the main brain of a standard computer, rather than a specialized graphics card often used for heavy AI work. After 24 of these training steps, the system reached a point where it could identify garbage with a certain level of accuracy. In the best moment of this run, the model correctly identified about 62 percent of the boxes it drew around objects, meaning it was usually right when it claimed to see something. However, it only found about 28 percent of the actual trash present in the images, missing many items that were clearly there. The overall score for how well the model matched the real objects was modest, reflecting the difficulty of the task and the small amount of data used.
Despite these limitations, the system proved it could work end-to-end. When shown a new picture of a lake, the model scanned the image in less than a second, drawing a box around each piece of trash it found and assigning a name to it. If the computer saw three boxes labeled "plastic bottle," it counted three bottles. This simple counting method, which relies on the detection step, showed that the pipeline is functional. The model was careful enough to avoid false alarms most of the time, but it was not yet sensitive enough to catch every single piece of floating waste. The low number of training images and the presence of confusing or placeholder labels in the data set likely contributed to the missed detections. Water itself is a difficult environment for vision, with glare, ripples, and reflections changing the appearance of objects, making it harder for the computer to distinguish a floating bottle from a patch of light.
The study concludes that while the current results are not perfect, the approach is viable for low-cost monitoring. The system runs on ordinary computer hardware, learns quickly, and successfully connects the steps of finding, naming, and counting trash. The researcher notes that the results are a snapshot of an unfinished project; the training was stopped before its full potential was reached, and the system has not yet been tested on a separate set of images it has never seen before. Future work will involve cleaning up the labels, adding thousands more images taken in different weather conditions, and testing larger, more powerful models. The ultimate aim is to deploy this technology on cameras attached to drones or boats, turning the slow, manual task of watching rivers into an automated, continuous stream of data that can help protect our waterways.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.