IMAGE BASED COFFEE BEAN CLASSIFICATION USING DEEP LEARNING TECHNIQUE.

Abstract

Coffee is one of the most famous gentle drink within the world. Most peoples take a cup of or double cup of coffee every morning to stimulate them even though they drink after the launch and totally, over 2.25 billion cups are fed on every day. Monetarily, coffee is the second most sent out commodity after oil, and utilizes more than 100 million individuals around the world. Coffee Arabica has developed for thousands of years in Ethiopia, in the southwestern highlands forests. The classification and grading of coffee in Ethiopian coffee quality inspection and certification center or Coffee board is manual. This leads to so many problems like pruning of error, inefficient, require a lot of labor and is not cost effective. This research was conducted with the objective of developing an appropriate computer routine algorithm that can characterize different varieties of coffee based on their growing region. To address this problem, we proposed deep learning approach for classification of coffee bean based on their growing regions. The proposed system has two main components namely, training model and testing the trained model or developing web application using flask. In this study, we applied different image preprocessing techniques such as: - removing noise, normalizing images and resizing images. We proposed two novel segmentation algorithms which are used to extract region of interest (ROI) and both are achieved excellent accuracy in this study. Two frameworks are proposed: one is to train a deep neural network model from scratch, and the other is to transfer learning a pre-trained network model. The model with best performance were obtained by testing different network layers, optimizers, learning rates, loss functions, number of epoch, batch sizes, and steps. Based on the optimized model. The developed classification model trained on 3120 dataset’s collected from ECQIAC. To increase the dataset, we applied different augmentation techniques. We split the dataset into different test options such as: - 90:10, 80:20 and 70:30 for training and testing respectively. The model classified the input coffee bean image into Gujji, Jimma, Kaffa, Nekempti, Sidamo and Yirgacheffe using softmax classifier function with 97.8 % accuracy. The entire system has been evaluated by employee from ECQIAC, the analysis shows that the system is effective to classify green coffee bean based on their growing region.