| Abstract: | Most agriculture products are main source of food and industries input, that have big contribution on human being day to day life activities. From most known agricultural products, haricot beans are popular and know edible leguminous product. In Ethiopian, this leguminous bean is the way to get income currency for growth of the country. It takes 15-21 % of market exchange in Ethiopian Commodity Exchanges organization. Haricot beans product need classification based on the level of its quality. Nowadays, the process of identifying quality of haricot is done manually by general inspection and just by looking using naked eye. This process takes much more extra amount of time and the quality measurement has low accuracy because, it is subjective and depends on the condition of the person doing the tasks. As a result of which, controlling the quality of haricot beans is not effective and efficient. To solve this major problem the current study proposes haricot bean grade classification by using digital image processing.
Digital Image processing technology is an emerging and growing technology to resolve this kind of practical and physical problem. Each level of the research was held by using experimental method. As experimental tool, we have used MATLAB software. For the experiment, the researcher collect sample haricot beans from ECX laboratory and then capture with image quality of 3264 by 2448 pixels. The captured images have some noises that appear from camera and environmental setting. To remove this noise media filtering technique is applied. After the binirized images were segmented by watershed segmentation techniques, convolutional neural network is employed as feature extractor and classifiers. The researcher used Add-16 feature extractor algorithm from ResNet-50 package. To train the classifier we have used 300 training image set and 90 individual test set. Experimental result shows that the model achieves 90.0% grade classification of haricot bean, which is a promising result. After all, classification algorithm have error of 10% of from individual test. So the researcher recommend that to minimize the rate of error happen in classification. |
