Feature Extraction and Visual Classification Methods for Identifying the Quality of Local Food Based on Digital Images

Abstract

Local food quality is essential for ensuring food security and enhancing product competitiveness in the market. Despite its importance, quality assessment predominantly depends on manual visual inspection, which is inherently subjective and inconsistent. This reliance contributes to post-harvest losses and diminished consumer trust. This study aims to develop a food quality classification system based on digital images by employing the EfficientNetV2B0 deep learning architecture through a transfer learning approach. The dataset consists of eight classes from four major commodities, namely apples, bananas, tomatoes, and bitter gourds, each in fresh and non-fresh conditions. All images were pre-processed through resizing and normalization, followed by data augmentation to increase variability and mitigate overfitting. The model was trained using a lightweight configuration and evaluated with accuracy, precision, recall, F1-score, and confusion matrix metrics. The results show that the proposed model achieved an overall accuracy of 99.8 percent, with precision, recall, and F1-score values approaching perfection across most classes. These findings demonstrate that the combination of transfer learning and data augmentation is effective in distinguishing between fresh and non-fresh food products based on visual features. In practical terms, this system has potential to support farmers and small-medium enterprises in conducting more objective, efficient, and consistent quality inspections