Vision Based Grasp Type Selection in Hand Prosthesis Using Convolutional Neural Networks

Amputation can be due to several causes: traumas, blood vessel diseases, diabetes, blood clots, and so on. Amputation causes a few complications for amputees such as hand dominance, psychology, social challenges, endure host of discomfort etc. These amputees must rely on a hand prosthesis to move their hands and grip objects. Some are electrically powered, while others require mechanical movement to control the prosthesis using harnesses. Therefore, an automated vision-based grasp type selection controller for hand prostheses is suggested and designed in this study, mainly to aid the users and allow them to grasp an object easily without any effort. In this study, three pretrained Convolutional Neural Networks (VGG-16, Inception-V3, and DenseNet-201) are formulated through transfer learning to classify the types of grasps. The three grasp types classified in this study are palmar wrist pronated, palmar wrist neutral, and pinch. The architecture of pretrained models was tuned by replacing the three last layers with two dense/fully connected layers and a dropout/regularization layer. Regularization and dropout are two commonly used techniques to prevent overfitting. As a result, VGG-16 trained with both dropout and regularization layer achieves perfect 100% accuracy in classifying the objects to their respective classes, followed by DenseNet-201 and Inception-V3 trained with only regularization layer with 98.4% and 98.2% accuracy, respectively.