Bionic hand movement recognition using BioMotionNet with U3AR-Net segmentation and hybrid optimization

The accurate and real-time Bionic Hand Movement Recognition is very central to the growth of prosthetics and assistive robotics. But the present ones face the problem of noise being interfered with, segmentation errors, feature misclassification, and computation inefficiency. The proposed study allows BioMotionNet- a novel deep hybrid framework-to serve these problems with advanced signal filtering, segmentation, feature extraction, and optimization. The BioMotionNet constituted by five components: (1) a Gradient-Enhanced Bilateral Median Filter (GEBM) for robust denoising and normalization; (2) U3AR-Net-a UNet3++ with Relaxed Stretch Fusion and enhanced Atrous Spatial Pyramid Pooling-for fine segmentation; (3) Histogram-Binary ResNet (HBR-Net) for hybrid feature encoding; (4) A selection algorithm based on mutual information called Dingo-Arithmetic Hybrid Optimization Algorithm (MI-DAHOA) for optimal selection; and (5) BioMotionNet-classifier for high-level motion recognition integrating DenseNet, DarkNet, and MHCNN-LSTM. The framework proposed attained an accuracy of 99.796%, F1-score of 99.880%, sensitivity of 98.175%, and specificity of 99.122%, along with a very low FPR and FNR, respectively. The BioMotionNet has outperformed other models in terms of faster convergence, better generalization, and robustness. BioMotionNet stands as a dependable noise-resilient and computationally economical solution for the Bionic Hand Movement Recognition process. Being hybrid in architecture, it surpasses the capabilities of traditional models and is probably suitable for practical deployment in prosthesis control, rehabilitation, and assistive technology.