Graph Neural Networks for Image-Based Quality Assessment of Perishable Goods

The state of perishable products is crucial for consumer health, nutritional quality, and efficient operations along the supply chain. Small visual cues, like discoloration and changes in texture, often reveal the overall condition of these products. This paper presents a graph neural network (GNN) framework for assessing the quality of perishable goods based on images. It combines texture analysis across different channels with relational learning. Traditional methods for classifying images often rely on the overall color mean or convolutional models. This technique, however, uses inter-channel co-occurrence features. In other words, it utilizes the Cross-Channel Gray-Level Co-occurrence Matrix to identify variations in chromatic texture that indicate spoilage. Input images go through preprocessing with adaptive color normalization in CIE Lab space. This process helps standardize lighting conditions while preserving natural hue and saturation patterns, which are crucial for distinguishing freshness. Following this, image-level graphs are constructed, and GNN-based classification categorizes items into pure-fresh, medium-fresh, and rotten states. The proposed method delivers high accuracy in classifying fruit and vegetable freshness into three categories. It outperforms CNN- and transformer-based baselines on a dataset of 60,000 images by managing intermediate freshness states that CNN-only approaches often miss. This solution is easy to expand and does not require invasive methods. It supports automated monitoring in retail, logistics, and consumer applications. It also connects with current trends in digital image processing for evaluating food quality.