Deep Learning Technique in Recognizing Hand Grasps using FMG signals

Grasping objects are common phenomenon in daily human activities. Force myography (FMG) signal, a noninvasive technique can record muscle movements while a human participant grasps different objects and be categorized using machine learning (ML) algorithms. In this paper, a popular deep learning technique is presented for hand grasp recognition. A novel convolutional neural network (CNN) architecture was implemented in learning grasps via force myography. Twelve participants wearing an FMG band on dominant hand's forearm performed six hand grasps. Training dataset consisted of one-handed grasping small objects of different shapes and sizes either wrapping or pinching with fingers with a variety of arm poses. The proposed FMG-based CNN model obtained cross-trial classification accuracy of 96% (population mean) and was found comparable with other ML techniques. Pretranined Alexnet (with ImageNet dataset) through transfer learning was implemented to classify the hand grasps for comparison. The proposed model outperformed the pretrained Alexnet in terms of validation accuracy, loss, and training time. For future FMG-based practical applications, it would be advantageous to use the model for transfer learning where comparatively smaller datasets are desirable for training purpose.