Prosthetic Hand Grasp Recognition Using Residual-Unet Combined with Cross Over Attention

Grasp pattern recognition plays a pivotal role in empowering prosthetic hands to accurate and effective grip categorization is the foundation of prosthetic hand functionality, therefore facilitating better user experiences and usability. This work presents a custom-built U-Net architecture improved with residual blocks and cross-over attention methods, especially designed for grip categorization tasks. While cross-over attention improves spatial and contextual learning across network layers, the proposed model uses the natural capabilities of U-Net for feature extraction reinforced by residual connections to reduce vanishing gradient problems and increase convergence. Two separate classification scenarios—Whole Category Classification (WCC) and Binary Outcome Classification (BOC)—were followed in experimental assessments. Under WCC, the model attained an accuracy of $83 \%$ and under BOC, $79 \%$ to show its dependability and flexibility throughout several categorization needs. This study highlights how sophisticated neural network designs might help prosthetic technology to progress and open the path for more responsive and easy artificial limb control systems