LoRA-Based Diffusion Data Augmentation for Underwater Garbage Detection: An Empirical Study with YOLOv8s

The ocean floor is a place where light struggles to reach, and where the water itself acts like a thick, hazy curtain. In these dim, murky depths, the plastic bottles, fishing nets, and discarded tires that humans have lost do not sit clearly against the sand. They are often distorted by the way water bends light, obscured by floating particles, or hidden in shadows. For scientists trying to clean up the oceans, this visual confusion is a major obstacle. To find and count this underwater garbage, they rely on cameras and computers, but the computers need to be trained on thousands of clear examples to learn what they are looking for. The problem is that clear underwater photos are incredibly hard to take. The environment is difficult to access, the equipment is expensive, and the images that do exist are often too blurry or dark to be useful for teaching a computer.

To solve this shortage of good training data, researchers have turned to a new kind of tool: artificial intelligence that can imagine and create pictures. These tools, known as generative models, can produce images that look like photographs without a camera ever being used. However, simply asking a computer to "draw a plastic bag underwater" often results in a picture that looks nothing like the real thing. The generated object might have the wrong shape, the wrong color, or appear in a place where it doesn't belong. If a computer learns from these fake images, it might get confused and fail to spot real trash later. The challenge, then, is not just making more pictures, but making pictures that are so convincing and accurate that a computer can learn from them as if they were real.

A team of researchers set out to test whether this approach could actually work for cleaning up the ocean. They focused on a specific type of artificial intelligence system capable of generating images, combined with a method to teach that system exactly what different types of underwater trash look like. They did not try to replace the need for real photos entirely; instead, they wanted to see if adding a small number of carefully crafted fake images to a limited set of real ones could help a computer detector become smarter. They chose to study five specific types of common debris: plastic bags, gloves, face masks, fishing nets, and cans. These items are notoriously difficult to spot because they are often transparent, crumpled, or blend in with the seabed.

The researchers began with a collection of real underwater photos that had already been labeled by humans. To ensure their test was fair and their results were trustworthy, they carefully separated these real photos into three groups before they started any computer generation. One group was used to teach the image-making system what the trash should look like. Another group was used to train the computer detector. The third group, which was never shown to the image maker or the detector during the learning phase, was saved strictly for the final test. This strict separation prevented the computer from using the test answers during the learning phase.

To make the fake images useful, the team used two special techniques. First, they taught the image generator to pay close attention to the specific details of each trash type. They showed the system a few dozen real examples of a plastic bag, a glove, or a net, and let the system learn the unique way these items look underwater—their folds, their transparency, and how the blue-green light of the deep sea affects their color. Second, they gave the system a simple map to follow. Before generating an image, the researchers drew a box on a blank canvas where they wanted the trash to appear. The system was then forced to place the generated object exactly inside that box. This ensured that the fake images would match the labels the computer detector expected to see, preventing confusion about where the object actually was.

The team created hundreds of synthetic images, but they were very selective about which ones they used. They manually checked the generated pictures, discarding any that looked too strange, had broken shapes, or did not match the label. Only the best images were added to the training set for the computer detector; specifically, just 50 additional synthetic images were included for each experimental setting. They then ran three different experiments to see what happened. In the first, the detector was trained only on real photos. In the second, they added fake images that were generated without any special training on the specific trash types. In the third, they added the fake images that had been carefully tuned to look like the specific trash items.

The results showed that simply adding more fake pictures was not enough. When the detector was trained with the un-tuned fake images, its performance actually got slightly worse. It became less accurate at finding trash and missed more items than before. This suggested that if the fake images looked too different from reality, they only served to confuse the computer. However, when the detector was trained with the carefully tuned images, the results improved. The computer became better at identifying the trash, finding more of it, and locating it more precisely. The accuracy of the detector rose from about 83 percent to 85 percent, and its ability to find the correct items improved as well.

The study also revealed that this improvement did not happen equally for every type of trash. The computer got significantly better at spotting gloves, cans, and fishing nets when it learned from the tuned fake images. However, for face masks and plastic bags, the improvement was much smaller, and in some cases, the performance did not change at all. This suggests that the method works best for objects that have a clear shape and texture, but it struggles with items that are very thin, transparent, or easily deformed. The researchers concluded that while these AI-generated images can help, they are not a perfect substitute for real data. They are a useful supplement that can boost performance when real photos are scarce, but they must be created with great care and checked for quality.

Ultimately, the work demonstrates that we can use artificial intelligence to help solve the problem of ocean pollution, but the technology must be guided by human understanding. The computer cannot just be told to "make more pictures"; it must be taught the specific visual language of the underwater world. By combining real data with high-quality, carefully controlled synthetic images, researchers can build better tools to monitor and protect the ocean floor. The study confirms that this approach is a promising step forward, offering a way to make the most of limited resources while acknowledging that the complexity of the underwater environment still requires careful handling.