DDABNet: a dense Do-conv residual network with multisupervision and mixed attention for image deblurring

Not all pixels are appropriate for image deblurring by traditional image deblurring methods. Similarly, not all feature information is appropriate for network model learning. To extract feature information effectively and ensure that the network pays attention to the feature information that is conducive to image deblurring, a dense Do-conv residual network with multisupervision and mixed attention is proposed, which is called DDABNet. First, we construct the Do-CA-Conv residual block by focusing on the important channels at the convolution level and densely connect the Do-CA-Conv residual blocks by summation in the first half and concatenation in the second half to construct DoDense(S &C) block. Second, at the third-scale stage of the encoder and decoder, to build interdimensional dependencies in the DoDense(S &C) block, we rotate the feature cube along certain dimensions to obtain an attention map and rotate it back before the first and second Do-CA-Conv residual blocks. Third, after the DoDense(S &C) block, followed by adaptive mixed-channel attention and spatial attention, which enhance attention to important information, the DoDense(S &C) attention block(DDAB) is constructed. Finally, the encoder and decoder outputs are constrained with multisupervision in the frequency domain. A visual inspection and the objective evaluation index of the benchmark datasets show the effectiveness of our proposed network.