<![CDATA[Manually classifying good quality coffee beans is subjective, can take a considerable amount of time and is hard to standardize among various operators. For this research, computer vision was used to create a classifying system, dividing coffee bean pictures into defect and Good Quality. Based on mobile execution performance, we evaluated five existing computer vision transfer learning models. Image datasets include those used to create it, public online collections of coffee bean images, a primary set collected for use during research and the full combined 1,102-image collection broken down into a training (858 pictures), validation (114 pictures) and test (130 pictures) set. We made images of the same 224x224 resolution, then used an augmentation pipeline that rotated, flipped randomly horizontally and added color changes to increase robustness. Normalized the pixel intensities using statistics gathered for ImageNet. Models were trained identically and used Adam for optimizer and a learning rate, batch size and epoch quantity. Densely Connected Convolutional Neural Network (DenseNet121), EfficientNetB0, MobileNetV2, Residual Network (ResNet50) and Xception all performed at equal settings during experimentation. The top accuracy level came from EfficientNetB0 and Xception, both reaching 96.92% on the testing data. We selected EfficientNetB0 as our core model for its performance, small size and steady use on a smartphone (as seen in the application prototype we made), but it was still a solid performing alternative to Xception. The Android prototype that came from our study supported photo input using either a camera or file upload to provide instant quality status. Transfer learning's ability to enable the use of models capable of automated and consistent assessment for coffee bean quality control would be an improvement in small coffee operations.]]>
Copyrights © 2026
