Lightweight Models for Fruit & Vegetable Defect Detection and Edge Deployment

With rapid growth of e-commerce, ensuring high standards of freshness of perishable items like fruits and vegetables is crucial for customer satisfaction. Wherein automatically detecting rotten and fresh fruits on mobile and edge devices poses a challenge. To address this challenge, we have explored lightweight and efficient deep learning models for fruit detection. This paper does a comparative study on fruit defect detection based of performance metrics like Accuracy, Precision, Recall and F1-Score. We compare LWCNN and three ImageNet-pretrained benchmarks-ResNet18, EfficientNetB0, and MobileNetV2 using accuracy, precision, recall, F1-score, model size, inference time, and parameter count. Among these ResNet18 provides best accuracy ranging between 97.77% to 100 % but has a slow inference time, larger size and a high number of parameters increasing the complexity. In contrast the LWCNN provides acceptable results ranging between 93.30 % 99.35 % accuracy, at a significantly reduced inference time, model size and number of trainable parameters, making it suitable for deployment in speed - driven scenarios such as quick commerce platforms.