Abstract
Man is the primary source of marine pollution, which is extremely harmful to both humans and the marine ecosystem. As a result of these wastes, humans may encounter difficulties when engaging in marine activities like fishing and gathering many other beneficial resources from the ocean. The initial point where the litter or trash starts is on the beaches, and then it ends up in the ocean. Plastic (bottles, bags, and caps), aluminium (cans, pull tabs), glass (bottles), and cigarette buts make up most of the trash discovered on beaches.
Information is crucial to understanding what is gathered and where it is collected and motivating companies to improve their methods and goods. After beach clean-ups, trash is sorted and counted to accomplish this. This is a laborious and time-consuming task. This study aimed to create a method for using images to recognize and count specific trash-related things.
Machine Learning and Deep Learning approaches, which are branches of Artificial intelligence, are the best techniques for object detection. However, the YOLOv4 (You Only Look Once) algorithm, which falls under the Deep Learning approach, was used for object detection. In this research, a new YOLOv4 algorithm was created that was a modification of the old one. The modification that was done was that 15 layers were removed from the neural network of the old one, and the new algorithm is now much faster without affecting the accuracy of the detection.
There were a couple of experiments done using the new YOLOv4 algorithm and the old YOLOv4 algorithm to detect objects such as cans and cigarettes buts from an image. The new YOLOv4 algorithm contains 147 layers in the neural network, and the old one contains 162 layers in the neural network. According to the testing results, the newly constructed neural network, which has 147 layers compared to the old neural network's 162 layers, operates more quickly with little decrease in efficiency. The system can detect a single can with an average accuracy of 98% for a new can and 89% for a faded can, according to previous studies. Additionally, the technology is only capable of detecting 20 cans per image. The accuracy for a single cigarette buts was about 87%, which is somewhat less than the cans, in subsequent studies using cigarette buts.
It can be concluded that using the new algorithm that was created, it goes through multiple images and calculates the total number of cans from them, so the objective is fully met in this research.