An algorithmic approach for automating the sorting and grading harvested TJC mangoes

Abstract

Abstract Mango grading is a quality assessment method carried out by mango exporting industries. Precisely graded mangoes support to uplift the demand of the industry as well as the revenue. Currently, both manual and machinery-based approaches are being used in the global market. Sri Lanka follows a manual grading approach to grade TJC mangoes which is the predominant exporting mango variety in Sri Lanka with a beautiful golden orange and unblemished skin. However, manual grading of mangoes using visual perception is laborious, inaccurate, and inconsistent. Therefore, this project aims to overcome these issues by introducing a machine learning and computer-vision based solution to grade harvested TJC mangoes in terms of their surface spots, size, and weight. This research has focused on the six perspectives of each mango to get an accurate final result from the machine learning model for the mango grade. This work is carried out in four phases: The first phase is the acquisition of images for training and testing purposes. Images were acquired by capturing six images per mango from six perspectives. The generated dataset consists of 1500 color images from five different quality classes. For experimentation purposes, the dataset was split into two groups containing 1350 and 150 images respectively. The larger group was used to train the classifier whilst the smaller group was used as the test dataset. Several image pre-processing techniques were followed in the second phase to enhance the images and extracted the essential features for the classification processes to identify the color scale, thresholding, noise removal, and contour detection techniques. Morphological operations and histogram of gradient (HOG) and local binary pattern (LBP) were used in the feature extraction phase. K-means clustering and K-medoid was used to detect the surface around the mango stalk. Sequential forward selection algorithm was applied to identify the best feature subset from the extracted superset of features in the feature extraction phase. The final grade classifier considered all the six sides of images in each mango. In the fourth phase, a model consisting of two main classifiers for analyzing the size and surface spots was implemented.