Abstract
Image restoration is an extensively studied area with lots of outstanding algorithms developed. Nevertheless, most existing methods still have some limitations that only apply to a single tailored restoration task or suffer from long iterative reconstruction time or yield unstable results. To address these challenges, this work presents a multi-noise and multi-channel enhanced Deep Mean-Shift Prior (MEDMSP) for grayscale IR tasks. Specifically, we draw valuable high-dimensional prior knowledge by learning a multi-noise stimulated DMSP network from color images with RGB-channels. Variable augmentation technique is then adopted for incorporating the higher-dimensional network prior into the iterative reconstruction procedure. MEDMSP has been evaluated on different IR tasks and compared to a variety of state-of-the-art methods. Experimental results show that the proposed method has better capability in image deblurring and accurate compressive sensing reconstructions in terms of both visual and quantitative comparisons.
Highlights
Image restoration (IR) is one of the most fundamental and popular topics in computational imaging
Deep learning has shown great potential in computer vision [17], [18] and it becomes a popular choice to solve IR tasks [19]–[22]. These methods can be regarded as the discriminative learning approaches, which directly learn the parameters in prior term prior(u) by optimizing a loss function on a number of clear-degraded image pairs
We present a multi-noise and multi-channel enhanced prior dubbed as multi-channel enhanced Deep Mean-Shift Prior (MEDMSP) for grayscale IR tasks, which introduces two new key features for deep mean-shift prior (DMSP) and a new iterative algorithm for incorporating high-dimensional prior into lower-dimensional task
Summary
Image restoration (IR) is one of the most fundamental and popular topics in computational imaging. A remaining drawback of DnCNN is its requirement for an expensive retraining whenever the scenarios (like the noise level, noise type or desired measure of fidelity) change a little bit To address this issue, a denoising autoencoder priors (DAEP) was proposed for different IR tasks [29], [31]. A deep mean-shift prior (DMSP) being proportional to the gradient of the logarithm of the image prior was proposed in [30] Both DAEP and DMSP train only one network and integrate it into the iterative restoration for different IR tasks. To enable MEDMSP more robust and stable for different IR tasks, multi-noise weighted strategy is adopted This design is motivated by the aggregation principle, indicating that the multi-noise stimulation can avoid getting into local solutions and make the iterative process more stable [16]. The rationality of setting the number of the multi-noise implementations in Eq (11) to be three will be verified in the Experiment Section
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
