An Improved Deblurring Method Based on DeblurGAN
Wenlong Yan, Gong Zhang, Weiwei Li, Wenbo Liu · 2025
Addressing the limitations of traditional generative adversarial networks in blurred image deblurring, such as the absence of edge details, subpar deblurring quality, and the chessboard effect, this paper introduces an enhanced deblurring approach based on DeblurGAN. The method integrates an Efficient Channel Attention module ahead of the convolutional layers within the generator's backbone network, modifies the deep residual network structure, and replaces the original network to enhance the model's capability to prioritize different image channels, thereby improving image restoration performance. An interpolation-plus-convolution technique is employed in place of transposed convolution during the generator's upsampling process to mitigate the fence effect in the deblurred images. In comparison to the original method, the improved algorithm demonstrates a significant enhancement in image deblurring quality.