Multiscale Image Deblurring Network Using Dual Attention Mechanism
Tao Zhang, Kerong Gai, Huihui Bai · 2022
The existing image de-blurring network based on deep learning is complex in structure, which has a large number of parameters. How to improve the efficiency of the network model without adding too many parameters is a key research direction. In this paper, a novel Dual Attention mechanism is proposed to improve the performance of Multiscale Image Deblurring Network (DA-MIDN). Here, dual attention mechanism is designed by combining spatial and channel attention to solve the problem that the convolution operation is limited by the size of the convolution kernel and the correlation between remote pixel features cannot be obtained. Specifically, the spatial attention module can make good use of the dependency between the distant pixels and capture the region of interest in the image information. Channel attention module calculates the correlation between the channel dimensions of the feature maps and suppresses the useless features, which can improve the training efficiency of the model. The experimental results show that the proposed method can achieve better de-blurring effect with a small number of parameters.