A Deep Motion Deblurring Network Using Channel Adaptive Residual Module

Abstract

In this paper, we solve the problem of dynamic scenes deblurring with motion blur. Restoration of images in the presence of motion blur necessitates a network design that the receptive field can completely cover all areas that need to be deblurred, while the existing network increases the receptive field by continuously stacking the ordinary convolutional layer or increasing the size of the convolution kernel. However, these methods inevitably increase the computational burden of the network. We propose a novel architecture consisting of a channel adaptive residual module. Different features of the blurred image are extracted and distributed on each feature channel. Our network can calculate the weight of each channel through learning, and extract the image features adaptively according to different degrees of blurring and importance of information. We embed the module in a modified encoder-decoder design with skip connections to achieve multi-scale feature fusion for further performance improvement. The extensive comparison with the existing techniques in the baseline dynamic scene deblurring dataset shows that the proposed network can effectively realize image deblurring, and the accuracy and speed are comparable with the existing techniques.

High-Frequency Attention Residual GAN Network for Blind Motion Deblurring

“…Currently GAN network deblurring is still the most popular deblurring network framework and many researchers have further improved the model on the basis of DeblurGAN. ChenY [32] proposed CARGAN and introduced a channel attention mechanism in GAN to extract the difference between channels adaptively. The quality of the deblurring results has been improved effectively.…”

“…Kubin et al [9] used some feature extractors in a modified WGAN to obtain perceptual loss, which achieved good results in the deblurring task. ChenY et al [17] also used a GAN network for the deblurring task and the effect was remarkable. The above studies show that GAN networks perform well on deblurring tasks.…”

“…Channel attention can adaptively adjust the weights of each channel of the feature map to improve the applicability of the network to different images. ChenY [17] also introduced a channel attention mechanism in the network to learn channel information that is conducive to image deblurring, and adaptively adjusts the weight of the channel. Spatial attention mechanism can deal with the blurring effect of different regions by using the spatial features of the image.…”

“…While the learning rate of the generator is set to after 50 epochs, with the initial parameter selected as . [17] and ChoS et al(2021) [21]. It is to be noted that the attention-adaptive and deformable convolution modules (AAM, DCM) are proposed in ChenL's method, which are considered to be used in methods such as DebluurGAN, MSCNN and SRN, etc in order to verify the effect of these two modules.…”

“…However, with the further increase in the number of parameters, the improvement of the restoration quality is limited while the model training time increases rapidly. Accordingly, the attention mechanism [16,17] is introduced into the model to alleviate the problem and further improve the training speed. However, as the network learns parameters based on the attention mechanism blindly, a lot of redundant information is learned, resulting in that more edge detail information of the restored image is missing.…”