Cascades of Regression Tree Fields for Image Restoration

Abstract

Conditional random fields (CRFs) are popular discriminative models for computer vision and have been successfully applied in the domain of image restoration, especially to image denoising. For image deblurring, however, discriminative approaches have been mostly lacking. We posit two reasons for this: First, the blur kernel is often only known at test time, requiring any discriminative approach to cope with considerable variability. Second, given this variability it is quite difficult to construct suitable features for discriminative prediction. To address these challenges we first show a connection between common half-quadratic inference for generative image priors and Gaussian CRFs. Based on this analysis, we then propose a cascade model for image restoration that consists of a Gaussian CRF at each stage. Each stage of our cascade is semi-parametric, i.e., it depends on the instance-specific parameters of the restoration problem, such as the blur kernel. We train our model by loss minimization with synthetically generated training data. Our experiments show that when applied to non-blind image deblurring, the proposed approach is efficient and yields state-of-the-art restoration quality on images corrupted with synthetic and real blur. Moreover, we demonstrate its suitability for image denoising, where we achieve competitive results for grayscale and color images.

Shrinkage Fields for Effective Image Restoration

“…The results in Tab. 1 show 7 that we outperform [2,3], and can also compete with the RTF-based cascade model [24] (trained with non-quantized images), whose additional flexibility does not seem pay off here since the image noise is truly Gaussian. The results further show that we can also compete for noise level σ = 15, for which we trained additional models.…”

“…The results in Tab. 2 show that a (5-stage) cascade of regression tree fields (RTFs) [24] achieves the best performance (trained with the same data as our models). This is not surprising, since the more flexible RTFs do not make any noise assumption (in contrast to all other approaches in Tab.…”

“…3 shows that our model scales to image sizes of more than 16 megapixels at reasonable runtimes (at most 10 minutes for our best model with a simple single-threaded Matlab implementation, and only 23 seconds on a GPU). While a cascade of RTFs [24] is very flexible and yields state-of-the-art restoration results, its relatively complex and highly optimized C++ implementation hinges on multithreading to boost runtime performance. Comparing singlethreaded performance (Fig.…”

Learning Collaborative Generation Correction Modules for Blind Image Deblurring and Beyond

“…For example, some works adjust the trade-off parameters [7,31] or iteratively change the prior terms [49] based on experiences to manually control the optimization process to avoid trivial solutions. Very recently, the learnable strategies [5,17,25,26,36,45,46] have also been introduced to help estimate the sharp image structures. However, both the manually designed tricks and trained networks will break the convergence guarantees of the standard optimization schemes.…”

Generic 3D Convolutional Fusion for Image Restoration

“…3 standard deviations σ ∈ {15, 25, 50} are chosen to measure the performance of 3DCF. Under such conditions, we compare our 3DCF with state-of-the-art DN methods as described in the introductory section 1: BM3D [8], LSSC [23], EPLL [42], opt-MRF [6], CRTF [29], WNNM [12], CSF [30], TRD [7], MLP [4], as well as the NN [5] fusion method. We use the same training data mentioned in [7], i.e.…”