A Mamba-Inspired Linear Vision Transformer with Spatial and Channel Reconstruction MLP for Roasted Coffee Bean Recognition

With rising consumer demands for coffee flavor and quality, quality control in coffee production processes has become increasingly important. Therefore, reducing labor costs during production while maintaining the consistency of coffee bean quality presents a significant challenge. To address this issue, this study proposes a Vision Transformer (ViT) architecture that incorporates a Spatial and Channel reconstruction Attention (SCA) block to assess the roasting levels of coffee beans. According to the experiment results, the proposed method achieves an F1-score of 99.79% on the roasted coffee bean database. Moreover, the method contains only 0.46 million parameters, making it suitable for deployment on lowcost embedded platforms.