ScholarWorks@성균관대학교

초록

The Widespread presence of plastics in modern life has led to a significant global challenge, such as plastic pollution. A critical aspect of this issue is the emergence of microplastics (MPs) from production waste; MPs are defined as plastic particles smaller than 5 millimeters. Those microscopic particles have penetrated the food chain, impacting both human and animal diets and raising serious concerns about potential health consequences. The current methods for detecting MPs primarily rely on analyzing tabulated chemical compound data or conducting time-consuming, costly laboratory tests. There is a remarkable gap in the application of machine learning and deep learning techniques, particularly through image analysis, for MP detection. Deep learning, especially transformer models like Vision Transformers (ViTs), offers such a promising solution for more accurate and rapid MPs identification and quantification. These models can automatically detect and quantify MPs by analyzing input images, offering more efficient alternatives to traditional approaches. Our proposed model addresses the inefficiencies and time constraints inherent in traditional methods. Using a Vision Transformer backbone within a Fast R-CNN model enhances the hybrid model's feature extraction and object detection capabilities. While this hybrid model shows strong performance in detecting particles of 3 mm or larger, in the current situation, it misidentifies very small particles (<1 mm). This is a crucial first step, cementing the way for future advancements in detecting even smaller particles and enabling real-time MP detection in vital water resources such as rivers, lakes, and oceans.

키워드