ENSEMBLE DEEP LEARNING FOR QUALITY CLASSIFICATION SYSTEM OF COFFEE BEANS BASED ON MULTI-COLOR WITH THE UTILIZATION OF SUPER-RESOLUTION BASED ON CONVOLUTIONAL NEURAL NETWORK FOR INPUT IMAGES

• Main Author: Norman Firdaus, Achmad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84262
• Institution: Institut Teknologi Bandung

Coffee is one of Indonesia's main commodities that significantly contributes to the country's foreign exchange. The quality of coffee beans affects both flavor and price, with quality standards referencing SNI 01-2907-2008. Currently, the quality assurance process for coffee is still manual, utilizing human visual organoleptic classification, which is time-consuming. To address this complexity, an automation system based on computer vision is necessary. Computer vision leverages visual information for decision-making and requires machine learning to recognize visual patterns. Deep learning, especially Convolutional Neural Networks (CNN), is effective in image feature extraction and can achieve coffee bean classification accuracy above 80%. The training procedure involves training data organized by certified assessors, with uniform lighting to avoid shadows. Training data is collected in a large container holding 1,053 coffee beans. Automatic segmentation techniques produce single coffee bean images, which are then cleaned of overlapping features. These single images have a resolution below the standard required by state-of-the-art (SOTA) deep learning, necessitating resolution enhancement through super-resolution. CNN-based super-resolution increases image resolution to enrich features, thereby enhancing the capabilities of SOTA deep learning models resulting from training. This improvement can be seen in the model's generalization and accuracy. Training involves four current deep learning models with seven color spaces (RGB, HSV, CIE LAB, XYZ, YCrCb, HLS, LUV) to better understand the defect features of coffee beans. The training results are combined with the soft-voting technique of ensemble learning, producing an accurate and robust model. Model evaluation using a confusion matrix shows 100% accuracy for super-resolved images using the FSRCNN (Fast Super-Resolution Convolutional Neural Network) method. This method addresses the ill-posed problem in resize interpolation methods, using convolution layers for feature extraction and deconvolution layers for resolution enlargement. The model remains accurate within the range of 99%- 100% even when the training data scheme is reduced to 30%, demonstrating good generalization. Keywords: FSRCNN, soft-voting, SOTA deep learning, color spaces, segmentation.