Design and Application of Hand Gesture Recognition in Embedded Image Processing System using Modified Convolutional Neural Network
Yiming Chen · 2025
Embedded image processing plays a significant role in analyzing visual data within compact system. Among the various applications, hand gesture recognition is an effective application for natural human-machine interaction specifically in sign language interpretation. However, recognizing sign language from hand gesture images is challenging task because of the high similarity among gesture images among different classes thereby reducing the recognition performance. Therefore, this research proposes a Gradient-Passing Quantizer based Convolutional Neural Network with Tunable Rectified Linear Unit (GPQ-CNN with TReLU) for hand gesture recognition in embedded image processing. This model integrates quantization-aware training with activation function thereby enabling to handle gradient flow during backpropagation while adjusting activation threshold. The ResNet50 is applied to extract discriminative features which captures both low-level and high-level spatial patterns thereby makes it better in identifying difference among similar gestures. During preprocessing, the median filter is utilized to effectively remove noise while preserving significant edge details that improves the gesture contour clarity. Following this, min-max normalization is used to scale the pixel values to uniform range among 0 and 1 which stabilizing the learning process. The proposed GPQ-CNN with TReLU achieves 99.96 accuracy and 99.63% precision on Indian Sign Language (ISL) dataset outperforming state-of-the-art techniques.