"Improving Mango Classification Accuracy with Enhanced Image Preprocessing and Neural Networks"

Abstract

This research introduces an automated system for mango grading using advanced image processing and machine learning techniques. Initially, classification on a raw dataset of 614 mango images, categorized into four quality grades (A, B, C, D), yielded an accuracy of 78%. The dataset was split into 70% for training, 15% for testing, and 15% for validation. To enhance accuracy, preprocessing steps such as resizing to 128x128 pixels, normalization (scaling pixel values to [0, 1]), and data augmentation (rotation, flipping, zooming) were applied. Additional techniques like color thresholding and blob detection highlighted features such as ripeness and defects. Reclassification after preprocessing improved accuracy from 78% to 92%. The VGG16 model was employed with transfer learning for robust feature extraction, using ReLU activation to capture complex patterns and a softmax layer for multi- class classification. MSE for Class A reduced from 0.015 to 0.009, and RMSE from 0.122 to 0.095, while Class D's MSE and RMSE also showed marked improvements. The confusion matrix revealed fewer misclassifications, validating the effectiveness of preprocessing and transfer learning in enhancing model performance. This system is a scalable solution for automated mango grading, offering potential benefits for agricultural quality control.