Multiscale Defect Extraction Neural Network for Green Coffee Bean Defects Detection

Abstract

With the increase in global demand for coffee, the issue of the health of drinks has attracted attention. Defective coffee beans contain substances harmful to health, emphasizing the importance of detecting defects in coffee beans. However, deep learning is rarely applied to the classification of multiple bean defects and can only distinguish high-quality from low-quality beans. This study proposed a multiscale defect-detection deep-learning model for coffee bean defect detection. In the architecture, multiscale fusion is used to improve the receptive field of each layer and thus obtain favorable defect features. Multiscale defect extraction is used to enable the extraction of coffee bean defect features of various scales. 7300 training and validation data were established for this study. The optimized model had the highest classification accuracy of 98.9% and the lowest of 84% for the types of defects, and the overall classification accuracy was 96%, higher than that of single-channel networks. The results revealed that the multiscale network can effectively extract and classify defects in coffee beans.

Keywords

• Coffee bean defect
• deep learning
• multi-scale feature fusion
• inception architecture
• CNN