Prediction of marine Litter based on marine information. This project was conducted during the Deeplearning Programming class in the first semester of 2024.
Marine litter threatens our living environment and ecosystems, emerging as a global environmental issue. It causes ecosystem destruction and economic losses, negatively impacting tourism and fisheries. Detecting marine litter in vast ocean areas and deep-sea environments is a challenging task. To overcome this, technology that can effectively detect and manage marine litter is needed.So, we chose marine litter prediction as our topic.
Dataset
We used marine litter images from the disaster safety data provided by AI-Hub.
This dataset consists of approximately 46,000 sonar survey images and about 18,000 underwater photography images. For this study, we only used the underwater photography images.
Preprocessing
- we removed those with multiple objects or no objects
- we set the ratio of train, validation, and test sets to 7:1:2.
- we used ImageNet normalization and resizing for preprocessing
Modeling
In this study, we used two models: YOLOv8n for the training model and SAM for pseudo labeling. For YOLOv8n, we used the model provided by Ultralytics, which was pre-trained on the COCO dataset and then fine-tuned on our dataset. We used AdamW (lr=0.001111, momentum=0.9) as the optimizer, set the batch size to 128, and ran a total of 50 epochs. For SAM, we used a pre-trained ViT-H as the backbone and performed segmentation using prompts from the bounding boxes generated by YOLO.
- The metrics for this experiment include precision, recall, and F1 score for class predictions, as well as mAP50 and mAP50-95 for bounding boxes. As shown in the table below, the results of YOLOv8s are higher in the test set compared to the train set, indicating that the model is not overfitting and demonstrates generalized performance. The results show that the test set is higher than the training set, indicating that the model is not overfitting and shows generalized performance.
- By using YOLO and SAM to effectively detect marine litter and obtain pseudo segmentation labels, we can reduce data construction costs and contribute to weakly supervised learning.
- By estimating the size and density of marine waste through segmentation, it is possible to prioritize waste disposal and reduce search time.