Efficient Microplastic Detection in Edible Food Images Using the Enhanced YOLOv8 with CBAM Architecture

Abstract

Microplastic (MP) pollution in edible food products has become a substantial global health concern that requires easy, accurate, and automated detection methods. Majorly used detection methods, such as FTIR or manual microscopic analysis, are slow and labor-intensive and not suitable for high-throughput screening. To address this limitation, this study proposes an advanced deep learning-based detection framework based on the YOLOv8 architecture with the integration of the Convolutional Block Attention Module (CBAM). A custom dataset was prepared and annotated with optical microscope (OM) images of edible food containing different types of MP, such as fragments, threads, and beads. The experimental results validate that the proposed model is better in small-scale MP detection in terms of precision, recall, and mean Average Precision (mAP). Experimental evaluation demonstrates the effectiveness of the model, achieving an overall precision of 98.32%, a recall of 98.30%, and an mAP of 98.81% across all classes. The model performed best for thread detection with 0.985 mAP, followed by fragments (0.984 mAP) and beads (0.862 mAP). After CBAM integration, the inference process became computationally efficient (an average of 8.1 ms per image of processing speed). The model reported reaching a small deployment size of 52 MB, comprising 218 layers and 25.8M parameters, and was inferred at 78.7 GFLOPs, which is computationally efficient enough to enable MP detection and is best suited for deployment in quality control, laboratory analysis, and environmental screening applications.