A Gesture Segmentation Method Based on Domain Adaptation and Channel Attention Mechanism
Kun Xu, Chen Mingyao, Yuan Xu, Xiaoxuan Li · 2021
Aiming at solving semantic inconsistency of segmentation results of HGR-Net network, a new Channel Attention Block (newCAB) is proposed in this paper. The original CAB is modified by embedding the global max pooling on lower stage to enhance the spatial context. The newCAB not only supplements the spatial information but also enhances the semantic information. In order to improve the generalization performance of our hand gesture segmentation model, an adaptive domain adaptation method based on a style transfer is proposed. VGG model is employed to reconstruct the source domain image according to the total loss function which is calculated by content loss and style loss. The stylization degree parameter is adjusted according to the complexity and the illumination conditions of the source image. The experiments on the OUHANDS dataset show that the value of mIoU and MPA of our network are 0.9324 and 0.9667 respectively, which are 5.6 and 5.2 percentage points higher than those of HGR-Net. The experiments on the self-collected dataset show that the mIoU and MPA of those images after applying our domain adaptation method are 0.8031 and 0.8899 respectively, and have been increased by 33.42% and 35% averagely. The proposed method based on domain adaptation and channel attention mechanism can improve the intra-class inconsistency and generalization performance.