Image semantic segmentation algorithm combined with attention mechanism

Abstract

There are still some shortcomings in the mainstream semantic segmentation algorithm, such as the loss of small size target and the segmentation inaccuracy. This paper improved the algorithm based on the High-Resolution Net (HRNet). Firstly, the channel attention mechanism and the spatial attention mechanism were integrated, which leads to a complementary relationship between spatial and channel information of features, retain more effective feature information, and improve the accuracy of feature extraction. In view of the large number of parameters in HRNet network model and the time-consuming training, this paper uses depth-separable convolution to reduce the parameters of the model without affecting the segmentation accuracy. Since the model maintains a high resolution of the image and retains the spatial information of the feature map, the segmentation effect of the small size target has been significantly improved. The mIoU in the PASCAL VOC2012 enhanced dataset reached 80.39%, Which is improved by 1.21% compared with the original model.