2014/11/28 by Girma, Biniyam, Getachew Abebe, Abebe, Getachew +2
Agricultural and Biological Sciences · Chemistry · Medicine · #Coffee research and impacts #Smart Agriculture and AI #Spectroscopy and Chemometric Analyses
paper · doi:10.20372/eajs.v7i1.146
openalex publication_date 2014/11/28 · openalex created_date 2017/04/07 · openalex updated_date 2026/07/01
The Hararghe coffee is known as Harar A, B, C and D in terms of its quality and growing regions. Coffee quality is usually determined subjectively by visual inspection which is prone to errors and also labour intensive. This necessitates the use of alternative methods, which can reduce errors and give a more accurate and objective measurement. In this work, an automatic sorting system consisting of computer routine algorithm to extract features of Hararghe coffee beans (HCB) images and artificial neural network (ANN) to classify these features was developed. Classification of four predefined HCB categories (Harar A, Harar B, Harar C and Harar D) was done by taking a total of one hundred sixty images (forty for each predefined class). The classification was carried out using three major feature categories i.e. colour, texture, shape, and a combination of the two features. These features were used as an input to the artificial neural network for classification and their results were presented using a confusion matrix. The classification accuracy of the 160 images is summarized as follows; using shape and size features 100%, 92.5%, 100%, and 100%, respectively as Harar A, B, C and D with an overall classification performance of 98.1% (157 images were correctly classified). Using texture features 92.5%, 97.5%, 85% and 95%, respectively with an overall accuracy of 92.5% (148 images were correctly classified). Classification results using colour features 100%, 77.5%, 97.5% and 100% respectively as Harar A, B, C and D. By employing colour features, out of 160 images used for classification 150 (93.8%) were correctly classified while 10 (6.3%) were misclassified. Finally using the entire twenty three features as an input a classification accuracy of 100%, 100%, 100%, and 97.5% respectively was achieved. In this scenario a total of 159 (99.4%) images were correctly classified while only 1 image (0.6%) was wrongly classified. In conclusion though the accuracy of this technique is unarguably acceptable for industrial purpose, it has to be investigated further using good number of images for its repeatability for it to be used in the day to day routine application.