Spatially Adaptive Image Deblurring Module Based on Dilated Deformable Convolutions
Hea-Bin Yang, Sung‐Jin Cho, Sung-Jea Ko · 2023
In this paper, we propose an image deblurring module based on modified deformable convolutions (MDCs) to deal with spatially varying blur. The proposed method first reduces the number of channels of the input feature map. Then, we extract spatially adapted features by applying a MDC that uses two 1D convolution kernels for the deformable offsets. An additional MDC with dilation enlarges the receptive field and further strengthens features. We fully preserve previous features and propagate them using shortcut connection. Experimental results show that the model incorporating our modules outperforms the baseline model.