Hemodynamic Analysis for Touch Induced Object Recognition using Convolutional Neural Network

The prime objective of this work is to develop a notable methodology for recognizing various objects through touch perception of disabled individuals using Functional near Infrared Spectroscopy (fNIRs) analysis technique. First, the fNIRs data has been procured from twelve blind folded healthy subjects while they guessed mentally the shape of objects presented on their palm. Five geometrical objects have been utilized for the current experiment which includes cone, cube, pyramid, sphere and cylinder. The acquired data have been fed to nirsLAB software to detect the Broadmann areas involved in this perception oriented task. Next, the procured data have been normalized in the range [0, 1] and then common average referenced to eradicate the effects of motion artifacts. Then, this data have been filtered to remove unwanted noise components and finally fed to an Independent Component Analysis (ICA) module to reinstate the independent components of the obtained data set. Then, these preprocessed data are transferred to the proposed Convolutional Neural Network (CNN) to categorize five distinct class labels. Performance analysis undertaken confirms that the proposed classifier is able to distinctly categorize the five class labels with a very high precision level and also outperforms other state-of-art classifier techniques. Hence, the proposed technique can be effectively implemented to cater the daily needs of patients suffering from blindness and communication oriented disabilities like Usher’s syndrome, Autism and such like.