Implementasi Statistic Histogram dan Gray Level Co-Occurrence Matrix Pada Klasifikasi Tingkat Kematangan Buah Kopi Berdasarkan Citra Multispektral

Abstract

Over the last decade, multispectral imaging techniques have appeared which are believed to be better than conventional imaging techniques, especially in examining the quality of various agricultural commodities. This technique has been developed to identify the level of maturity of coffee cherries by taking images using a special camera. The camera has specifications in the form of a wide electromagnetic spectrum, controlled illumination space, and narrow LED bandwidth. The image produced by the camera has 15 color channels with a wavelength between 400-1000 nm. In one study, these multispectral images were processed using a Convolutional Neural Network (CNN) to extract high-dimensional patterns automatically. CNN's high complexity allows the model to capture more complex features but will require more time and computational resources for model training and testing. Therefore, the classification stages in this study were carried out using simpler methods such as Naïve Bayes and Support Vector Machine (SVM). In processing, the multispectral image of coffee beans needs to be segmented first to get the right image of the coffee fruit object. Segmentation combines several methods such as gaussian blur, Sobel edge detection, morphological operations such as erosion, dilation, and hole filling, bounding box creation, and selection using circle detection. The success rate of segmentation using this method reaches 90%. A well-segmented image is extracted from color and texture features using a Color Histogram and Gray Level Co-Occurrence Matrix (GLCM). The resulting features are as many as 60 color features and 240 texture features. Then these features are implemented in classification using Naïve Bayes and SVM. The experiment found that the combination of color and texture features has the best performance with an average accuracy rate of 93% for Naïve Bayes and 96% for SVM. However, the use of feature selection in classification cannot improve accuracy.