Comparison Performance on SOTA Deep Learning Models for Coffee Beans Grading Inspection

Several coffee producers rely on manual inspection for quality control, which is prone to inconsistencies and time consumption. Recent studies have proposed Comparison performance on State-of-The-Art (SOTA). Deep learning models were investigated as part of the proposed intelligent computer vision for faster and more reliable coffee bean grading. Performance Deep learning models, i.e., MobileNetV3Small, ResNet50, EfficientNetV2L, and DenseNet201, were evaluated to grade the quality of coffee beans that followed a grading procedure based on Indonesian National Standard (SNI) 01-2907-2008. A specific tray was designed to position coffee beans evenly without overlap for sample preparation, which consist of about 500–600 coffee beans. The process of capturing images is conducted systematically, focusing on each type of coffee bean defect individually. For comparison of SOTA models, five classes of coffee bean defects, With 1053 beans per class, the total dataset comprises 5265 beans. This dataset was divided into 10% for testing and 90% for training. Within the training data, 12% was reserved for validation. The result from this paper shows that MobileNetV3Small achieves the highest accuracy (99.43%) with minimal memory and parameters. Its efficient design allows for robust feature extraction, crucial for distinguishing subtle defect features. While larger models like ResNet50 and EfficientNetV2L also perform well, they exhibit higher memory and parameters.