Ripeness Classification of Cavendish Bananas using Multi-object Detection Approach

Cavendish bananas, recognized for their global economic and nutritional value, often pose challenges in ensuring consistent ripeness quality throughout the supply chain, which is essential for efficient post-harvest management and ensuring optimal consumer experience. Traditional methods, reliant on human visual inspection, are subjective and inconsistent. This paper proposes a novel multi-object detection approach for accurately classifying the ripeness of multiple Cavendish bananas in an image using computer vision and deep learning techniques. We employed the Yolov5 model, a state-of-the-art object detection architecture using a standardized ripeness classification system to simultaneously identify and categorize every banana in an image. To train the model, a diverse dataset of 600 Cavendish banana bunches at different ripeness stages was collected, augmented, and annotated with ground truth labels. Training result shows that the multi-object detection network successfully delineates the bananas in input images while simultaneously predicting their ripeness classes. The proposed approach achieves 98.8% mean average precision, 90.5% precision, and 92.6% recall, even when the test images contain overlapping bananas and diverse background colors. The compact size of the object detection model, made it applicable to an embedded system, enabling realtime ripeness classification of Cavendish bananas with simple image uploads from handheld devices. This research contributes to the advancement of agricultural technology and opens avenues for future studies in fruit ripeness analysis and food quality assessment.