Improved Green Coffee Bean Recognition by Concatenating Different Image Enhancement Methods to MobileViT

We developed a method to enhance the recognition accuracy of green coffee beans by combining different image enhancement techniques. In image classification, color and contours were used as main features for recognition with contours being more prominent for green coffee beans. To improve the model's recognition capability of MobileViT, we employed nine common image enhancement methods for feature extraction and model training. However, individually training various preprocessing algorithms required excessively long training times and a large total number of model parameters. Therefore, we replaced the original image's RGB channels with image processing methods that exhibited higher correlations with ground truth for model training. We performed feature extraction on the mentioned image enhancement methods and analyzed the correlation coefficients between the model's recognition results. We selected algorithms that met the predefined threshold, concatenating the results of the chosen three methods as the new inputs for the model. We proposed the classification method without any preprocessing as the baseline. Among various combinations, the combination of bit-plane slicing, histogram equalization, and unsharp masking achieved an accuracy of 96.9%, representing an improvement of approximately 5.5% compared to the original method.