Deep Convolutional Learning for Automated Multi-Class Quality Grading and Defect Recognition of Fruits and Crop Leaves

Grading harvested produce by eye is slow, depends heavily on the inspector, and yields inconsistent results from one batch to the next, which contributes to post-harvest loss and unfair pricing. This paper presents an automated grading framework based on a convolutional neural network (CNN) that goes beyond identifying which crop is in an image and instead judges its condition. Images of fruits and leaves are assigned to one of five quality grades—healthy, ripe, unripe, pest-damaged, and rotten—so that maturity stage and common defects are recognized within a single model. The pipeline combines image normalization and enhancement, extensive data augmentation, hierarchical feature learning, and a softmax classifier trained with cross-entropy loss and the Adam optimizer. On a held-out test set the model reaches 94.9% accuracy and a macro-averaged F1-score of 94.8%, surpassing a support vector machine trained on hand-crafted features by 8.4 percentage points while needing only about 21 ms per image. These results indicate that the approach is accurate and fast enough to support real-time sorting and supply-chain decisions in precision agriculture.