Optimized YOLOV8: An efficient underwater litter detection using deep learning

Abstract

Underwater litter has been a major issue in preserving the marine ecosystem. Human waste is deposited into lakes, rivers, and seas which leads to polluted water. The underwater litter harms aquatic life and pollutes water bodies and ecosystems. Therefore, there is a need for effective and efficient methods for detecting underwater litter. An improved YOLOV8s model is proposed for the detection of underwater litter. The fine-tuning of all the YOLOV8 variants was performed to choose the best model i.e. YOLOV8s. OFAT technique examines how various configurations affect the performance of the optimized YOLOV8s model. YOLOV8s was used to optimize and tune the hyperparameters of the model. Additionally, two hyperparameter tuning techniques were compared, and the results demonstrated that the OFAT is the superior optimization approach. Additionally, the research compares the underwater litter detection results of the optimized model and the pre-trained model of YOLOV8s. The “UW_Garbage_Debris_Dataset,” dataset comprises of 15 different classes of underwater litter which were used to train the dataset for the proposed research. From experimental results, the optimized YOLOV8s model showed an outstanding precision of 98.8 %. In comparison with other optimizers, learning rates, batch sizes, and epoch sizes, the optimized YOLOV8s performed better at 64 batch size in terms of effectiveness and efficiency. ICRA19 and UW_Garbage_Debris_Dataset were the two datasets used in the proposed study to test the proposed optimized YOLOV8s model. In terms of effectiveness and efficiency, the UW_Garbage_Debris_Dataset gave better results in comparison with the studies of literature. Furthermore, the research synthesis was conducted which shows the overall model’s performance is outstanding. Future research should try various optimizers, batch sizes, and learning rates, as well as other hyperparameters tuning techniques.