Authors: Shrikant Patne, Professor Bangar Abhishek Popat, Professor Bhosale Sachin Bajirao
Abstract: Food that looks fine on the shelf is not always safe to eat, yet the decision is still made, for the most part, by human inspectors working under time pressure. This paper presents a machine-learning pipeline that grades fresh produce into three quality classes (fresh, medium and spoiled) from ordinary RGB images. We compare three classical learners (a support vector machine, k-nearest neighbours and a random forest) built on hand-crafted colour and texture descriptors with three deep models: a lightweight custom convolutional network, ResNet50 and a fine-tuned MobileNetV2. On a balanced dataset of 7,450 images, the fine-tuned MobileNetV2 reached 96.7% accuracy and a macro F1-score of 96.7%, which is 7.6 percentage points above the best classical baseline, while staying small enough (2.6 million parameters) to run on a mid-range smartphone CPU in roughly 38 ms per image. Ablation experiments indicate that data augmentation and transfer learning account for most of the improvement, and Grad-CAM maps confirm that the network focuses on decay-related regions such as dark patches and wrinkled skin. We also discuss where the approach still struggles, mainly the borderline medium class, and outline a route toward sensor fusion for real-world deployment.
