Effect of Applying the YOLO Object Recognition System for Developed Lake Underwater Images Database

Abstract

There are differences between underwater physical particles at the seas and lakes, where lakes underwater tend to be more brownish or greenish because lakes are inland bodies of water which do not have direct contact with the seas. Lake underwater image databases with higher turbidity images are difficult to find. Thus, there is a need to create a dataset for this kind of images to be used for real lake underwater research. For application such as underwater robot, it a system is needed that can distinguish objects in the lake when looking for them. As for object selection for this database, the objects are selected based on the assumption that these kinds of objects may fall into the lake and there is a need to search and find them back. Therefore, a method that could recognize these kinds of objects are important in underwater searching process. The study's goal is to develop an image database for lake underwater images and investigate the accuracy of object recognition system in acquired different lake underwater conditions. The YOLOv3 has been used in this study as a method of identifying the object in the image. The total of 315 images are used, where the ratio is 80% for training and 20% for testing. The tools utilized in this study are the LabelImg and Google Colaboratory softwares. According to the result and analysis, when testing with all of the experiments under various lake underwater settings, YOLOv3 has achieved overall accuracy of 92.32% for given underwater conditions.