Hybrid Feature Fusion and XAI for Low-Cost Powdered Food Adulteration Detection

Abstract

Food adulteration in powdered food products is a critical issue that affects public health and food quality, particularly in regions where rapid and low-cost testing methods are limited. Traditional laboratory-based detection techniques are often time-consuming, expensive, and inaccessible for real-time consumer use. This research proposes a hybrid explainable computer vision framework called Computer Vision–Adulteration Fusion Network (CV-AFNet) for detecting adulteration in powdered commodities such as maida and red chilli powder using image-based analysis. A custom dataset is developed using mobile phone cameras under both controlled and real-world environmental conditions, incorporating multiple adulteration ratios to improve model robustness. The proposed framework integrates preprocessing and hybrid segmentation techniques to improve the quality of the extracted sample region. Both handcrafted features, including color moments and texture descriptors, and deep features extracted through a lightweight convolutional architecture are combined using a feature fusion strategy. An Enhanced Recursive Feature Elimination (ERFE) method is applied to remove redundant features and improve generalization performance while reducing computational complexity. Multiple classifiers are evaluated to validate the effectiveness of the fused feature representation. In addition, explainable artificial intelligence techniques are incorporated to improve transparency and interpretability of the model predictions. The system provides a cost-effective, non-destructive, and practical solution for consumer-level food adulteration detection and supports future development of real-time mobile-based food quality monitoring systems.