Coastal Litter Detection through Image Analysis-Employing Deep Learning to Identify Microplastics-

抄録

The challenge of coastal litter accumulation led to the creation of a detection system powered by deep learning, aimed at identifying microplastics. The system harnessed the yolov7 [1] deep learning architecture, known for its proficiency in real-time object detection, and integrated the SAHI (Slicing Aided Hyper Inference) [2] vision library to augment its capabilities. Within the scope of our study, we conducted four separate evaluations using two versions of yolov7—the base model and the advanced yolov7-e6e—alongside SAHI. The performance of each setup was measured against a set of metrics, including Intersection over Union (IoU), Precision, Recall, F-measure, and Detection Time, recorded in seconds. The dataset for the study was composed of images sourced from real-world beach clean-up sites, including Hokuto Mizukumi Park. The detection algorithm was subjected to 700 rounds of training, with an initial learning rate of 0.001. Our findings indicated that the system was adept at identifying relatively small microplastics.