Postharvest Grading Classification of Cavendish Banana Using Deep Learning and Tensorflow

Abstract

In the Philippines' banana export industry, the postharvest classification of Cavendish banana is an important factor which affects one of the country's revenue. The present set up uses a quality inspector and is solely dependent on its visual capability. This may result in an error and inconsistent results grading classification. This study explores the use of image processing with deep learning approach to classify Cavendish banana grade by the basis of its standard requirement. A total of 1116 bananas comprising of 279 images from each category such as Class A big-hand or small-hand, Class B big-hand or small-hand and Cluster class (part of hand) are used to develop algorithm and classification accuracy by using Phyton OpenCV and Tensorflow. The sample images follow four stages: Image Thresholding, Feature Extraction, Classification Prediction, and Testing. An analysis of predicted accuracy in all classes (A and B) are more significant in the finger size value extraction (small or big hand) than in surface defects value extraction. The final classification in all trained, validation, and test data showed above 90% accuracy in all four classes. Hence, the proposed CNN classification in Tensorflow model can be commercially developed as a field-based complete automatic postharvest classification system for grading a Cavendish banana.