Light-efficient channel attention in convolutional neural networks for tic recognition in the children with tic disorders

Abstract

Tic is a combination of a series of static facial and limb movements over a certain period in some children. However, due to the scarcity of tic disorder (TD) datasets, the existing work on tic recognition using deep learning does not work well. It is that spatial complexity and time-domain variability directly affect the accuracy of tic recognition. How to extract effective visual information for temporal and spatial expression and classification of tic movement is the key of tic recognition. We designed the slow-fast and light-efficient channel attention network (SFLCA-Net) to identify tic action. The whole network adopted two fast and slow branch subnetworks, and light-efficient channel attention (LCA) module, which was designed to solve the problem of insufficient complementarity of spatial-temporal channel information. The SFLCA-Net is verified on our TD dataset and the experimental results demonstrate the effectiveness of our method.