Explainable AI for sign language recognition models: Integrating Grad-Cam LIME and Integrated Gradients.

Sign language recognition is crucial in bridging the communication gaps between hearing and deaf communities. In this study, we build on an existing sign language classification model based on the VGG19 architecture, enhancing its robustness through dataset augmentation and alternative data representations. We introduce a segmentation-based approach that utilizes hand masks generated by the U-Net model, replacing depth images to mitigate noise and improve classification accuracy. We use an Explainable Artificial Intelligence (XAI) approach that incorporates Grad-CAM, LIME, and integrated gradients methods to interpret the model's decision-making process, ensuring transparency and reliability in real-world applications. Our comparative analysis between Red-Green-Blue (RGB) and mask-based models demonstrates that while the RGB model benefits from richer texture and color information, the mask-based model effectively focuses on hand shape and structure. The integration of XAI further validates our results by highlighting key regions of the image that influence the model's predictions and by enabling a multi-perspective analysis that captures complementary aspects of the input, including region-based attention, pixel-level attribution, and structural shape analysis, thereby facilitating a deeper understanding of the model's internal representations and potential failure modes. This work contributes to advancing American Sign Language recognition by improving model generalisation and explainability, ultimately fostering greater trust and usability in assistive technologies.