In intelligent healthcare monitoring, vision-based sensing offers a practical route for non-contact medical diagnostics, yet tongue-coating assessment in traditional Chinese medicine (TCM)-related clinical settings is often hindered by background interference, illumination variation, and limited interpretability. In this study, we present an explainable vision-sensing signal-processing framework built as a two-stage segmentation–classification pipeline. First, a U-Net model segments the tongue region from RGB tongue images to suppress non-tongue artifacts and standardize the sensing signal. Second, encoder features learned during segmentation are transferred to a feature-attention classifier to recognize four coating textures (thin, thick, peeled-like, and mirror-like). To support trustworthy clinical use, gradient-weighted class activation mapping (Grad-CAM) is employed to visualize discriminative regions behind predictions. Experiments demonstrate improved robustness and accuracy over a conventional VGG-based baseline, while providing interpretable evidence. The proposed method serves as a deployable component for vision-centric intelligent sensing and can be integrated into broader healthcare monitoring systems.