Revolutionizing Hand Gesture Recognition: A Transfer Learning Approach using Surface Electromyography and Convolutional Neural Networks

The popularity of deep learning algorithms has increased significantly over the past few years, which are unrivaled in their capacity to automatically discover discriminant features from massive datasets. Deep learning techniques are rarely used for electromyography-based gesture detection due to the large amount of work involved in creating thousands of instances. In order to achieve high generalization and minimal training burden in Surface Electromyography (SEMG)-based gesture recognition, this research provides an effective Transfer Learning (TL) technique. This study aims to determine whether or not a certain EMG signal, captured by a sensor-based band, corresponds to a particular hand gesture. In order to identify and categorize objects, surface EMG and Machine Learning (ML) methods are employed. Suitable pre-processing processes are applied to the raw EMG signal recorded by the sensor in order to remove noise artifacts. We use transfer learning and convolutional neural networks to discover a good controller for a myoelectric prosthetic hand. First, pictures are generated from the sEMG signal using the continuous wavelet transform (CWT), which is a robust feature for the separation of various hand gestures. Then, we used what we learned from many image classification neural nets, including AlexNet and ResNet-18, to sEMG classification. When these deep neural networks were applied, they beat standard machine learning methods regarding precision and quickness. State-of-the-art accuracy levels of 100% on the publically available datasets, for example, have been reached using a hybrid of CNN and VGG16 for these data sets. To obtain excellent accuracy in two sEMG datasets, this research establishes an algorithmic based on Continuous Wavelet Transform (CWT) and CNN-VGG16 deep neural networks.