Abstract
In image deblurring, information from the regions where the blur was propagated is needed for effective deblurring. For example, large blurs, such as those caused by fast-moving objects leaving a trail of afterimages, need spatial context from a large region, while small blurs, such as those caused by slight camera shake, need spatial context from a smaller scope. In this paper, we used multi-scale features to provide the spatial dependencies needed to deblur non-uniform blurs. Compared to previous works, we efficiently extract multi-scale features using two approaches: (1) coarse-to-fine scheme that can extract multi-scale features by applying the network to different scales of the images, and (2) dilated convolutions that can extract multi-scale features by using different dilation rates. Combining both methods has a multiplicative effect since multi-scale features from dilated convolutions are extracted from the input images at different scales (i.e. coarse-to-fine scheme). Furthermore, we optimized our network architecture by using parallel convolutions to decrease the execution time of the deblurring process. We show that our proposed method has better results than state-of-the-art methods in terms of image quality and execution time.
Highlights
A blurry image has missing or distorted information that could be useful for personal viewing or other image processing tasks
This definition assumes that the blur kernel is uniform for the entire image, and it fails if the blurs are non-uniform
Our method of generating multi-scale features is done using coarse-to-fine scheme and aggregated dilated convolution. Both methods can extract multi-scale features by themselves; combining the two methods have a multiplicative effect by extracting multi-scale features at different scales of the images
Summary
A blurry image has missing or distorted information that could be useful for personal viewing or other image processing tasks. The blurring process is defined as the convolution of a sharp image and a blur kernel to produce the blurry image so that the blur kernel is approximated rather than the entire sharp image [1]. This definition assumes that the blur kernel is uniform for the entire image, and it fails if the blurs are non-uniform (i.e. different from region to region). Uniform blurs could happen because of camera shake the entire image is affected. Non-uniform blurs are common in dynamic scenes where moving objects cause motion blurs.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have