Abstract:
Accurate evaluation of the quality of produce is essential for food safety, consumer
acceptance, and reduction of post-harvest losses. Manual inspection and chemical
tests, which are considered traditional methods, are subjective, tedious, and destructive.
The proposed approach is a Multi-Modal AI System that detects and
predicts the freshness and ripeness of produce through a non-destructive intelligent
system, helping to overcome these limitations. The methodology employs
a two-stage hierarchical pipeline that integrates visual information from a camera
module with data from a sensor array comprising MQ4 (methane), MQ135
(CO2/ammonia), TGS2602 (Volatile Organic Compounds), and an AS7263 Near-
Infrared (NIR) spectrometer. In Stage 1, an object detection model identifies
the type of produce. In Stage 2, visual features extracted by a Convolutional
Neural Network (CNN) are combined with normalized sensor-array readings in
a novel early fusion model to determine freshness and ripeness. The findings
clearly demonstrate the strength of data fusion, with the multi-modal framework
achieving an overall prediction accuracy of 95%. This approach outperformed the
unimodal methods, which achieved accuracies of 93% for image-only and 88% for
sensor-only data, respectively. Additionally, the Vision Transformer (ViT) model,
designed specifically to detect artificially ripened fruit, achieved accuracies of 94%
for bananas and 92% for mangoes. The system is deployed through three applications,
including a “GreenMate” physical device, an “AI Food Assistant” mobile
application, and a standalone portable device specifically designed for the classification
of artificial ripening in mangoes and bananas. Together, these solutions
provide a low-cost, real-time, and practical approach to enhancing supply chain
transparency, protecting consumers, and reducing food waste.
