Deep learning for extraneous material detection in a rice processing factory

Abstract

In recent years, rice-based fresh foods, such as triangular rice balls and bento meals sold in convenience stores, have become popular commodities, primarily supplied by rice processing factories. To ensure food safety, food manufacturers must implement monitoring processes from rice storage bins to processing and finished products. Beyond independent management and compliance with food safety laws, preventing the contamination of extraneous materials before shipment remains a critical issue. Currently, most rice materials and finished products undergo metal detection, whereas non-metallic foreign matter cannot be identified using conventional detectors and thus requires manual inspection. To address this challenge, this study employs image sensors to capture characteristic information from both food materials and contaminants, and constructs a deep learning model to facilitate data analysis and classification, thereby establishing an evaluation model for foreign matter contamination in rice. The study incorporates monitoring across multiple process checkpoints, employing 10 different CNN models with parameter adjustments to compare precision rates and enhance detection accuracy. The results showed that the ShuffleNetV2 model, when trained for 50 epochs, achieved the highest performance for foreign matter recognition, with an average F1-score of 0.971. When trained for 100 epochs, the same model reached an even higher average F1-score of 0.982, demonstrating its strong effectiveness in foreign matter detection. Based on the development and comparison of these models, this study aims to enhance the factory's ability to identify and classify foreign matter, thereby reducing the risk of food safety incidents.