การจำแนกระดับความสุกของกล้วยโดยใช้การประมวลผลภาพแบบหลายปริภูมิสีและการเรียนรู้เชิงลึกเพื่อประเมินความสามารถในการทำงานทั่วไปของโมเดล

Abstract

This research presents a banana ripeness classification system based on image processing and deep learning. The objectives were to investigate the effect of color representations on model performance and to evaluate model generalization under different real-world imaging conditions. Two datasets were used: the Banana Ripeness Classification Dataset and the Banana Images - Real Dataset. Banana ripeness was classified into four classes: unripe, ripe, overripe, and rotten. Three color representations, namely RGB, RGB+HSV, and Grayscale, were compared using Random Forest, K-Nearest Neighbors, Support Vector Machine, CustomCNN, and MobileNetV2. The results showed that RGB+HSV improved classification performance compared with RGB and Grayscale. CustomCNN with RGB+HSV achieved the highest accuracy of 98.40%, while MobileNetV2 demonstrated stable performance across several experimental conditions. However, the Cross-Dataset Evaluation showed a considerable decrease in accuracy when the models were tested on images with different illumination, background, and image quality. These findings indicate that color representation directly affects banana ripeness classification performance and that cross-environment evaluation is essential for developing practical smart agriculture systems.