CIE XYZ Net: Unprocessing Images for Low-Level Computer Vision Tasks.

Cameras currently allow access to two image states: (i) a minimally processed linear raw-RGB image state (i.e., raw sensor data); or (ii) a highly-processed nonlinear image state (e.g., sRGB). There are many computer vision tasks that work best with a linear image state, such as image deblurring and image dehazing. Unfortunately, the vast majority of images are saved in the nonlinear image state. Because of this, a number of methods have been proposed to "unprocess" nonlinear images back to a raw-RGB state. However, existing unprocessing methods have a drawback because raw-RGB images are sensor-specific. As a result, it is necessary to know which camera produced the sRGB output and use a method or network tailored for that sensor to properly unprocess it. This paper addresses this limitation by exploiting another camera image state that is not available as an output, but it is available inside the camera pipeline. In particular, cameras apply a colorimetric conversion step to convert the raw-RGB image to a device-independent space based on the CIE XYZ color space before they apply the nonlinear photo-finishing. Leveraging this canonical image state, we propose a deep learning framework, CIE XYZ Net, that can unprocess a nonlinear image back to the canonical CIE XYZ image. This image can then be processed by any low-level computer vision operator and re-rendered back to the nonlinear image. We demonstrate the usefulness of the CIE XYZ Net on several low-level vision tasks and show significant gains that can be obtained by this processing framework. Code and dataset are publicly available at https://github.com/mahmoudnafifi/CIE_XYZ_NET.

A Little Bit More: Bitplane-Wise Bit-Depth Recovery.

Imaging sensors digitize incoming scene light at a dynamic range of 10-12 bits (i.e., 1024-4096 tonal values). The sensor image is then processed onboard the camera and finally quantized to only 8 bits (i.e., 256 tonal values) to conform to prevailing encoding standards. There are a number of important applications, such as high-bit-depth displays and photo editing, where it is beneficial to recover the lost bit depth. Deep neural networks are effective at this bit-depth reconstruction task. Given the quantized low-bit-depth image as input, existing deep learning methods employ a single-shot approach that attempts to either (1) directly estimate the high-bit-depth image, or (2) directly estimate the residual between the high- and low-bit-depth images. In contrast, we propose a training and inference strategy that recovers the residual image bitplane-by-bitplane. Our bitplane-wise learning framework has the advantage of allowing for multiple levels of supervision during training and is able to obtain state-of-the-art results using a simple network architecture. We test our proposed method extensively on several image datasets and demonstrate an improvement from 0.5dB to 2.3dB PSNR over prior methods depending on the quantization level.

Learning Raw Image Reconstruction-Aware Deep Image Compressors.

Deep learning-based image compressors are actively being explored in an effort to supersede conventional image compression algorithms, such as JPEG. Conventional and deep learning-based compression algorithms focus on minimizing image fidelity errors in the nonlinear standard RGB (sRGB) color space. However, for many computer vision tasks, the sensor's linear raw-RGB image is desirable. Recent work has shown that the original raw-RGB image can be reconstructed using only small amounts of metadata embedded inside the JPEG image [1]. However, [1] relied on the conventional JPEG encoding that is unaware of the raw-RGB reconstruction task. In this paper, we examine the ability of deep image compressors to be "aware" of the additional objective of raw reconstruction. Towards this goal, we describe a general framework that enables deep networks targeting image compression to jointly consider both image fidelity errors and raw reconstruction errors. We describe this approach in two scenarios: (1) the network is trained from scratch using our proposed joint loss, and (2) a network originally trained only for sRGB fidelity loss is later fine-tuned to incorporate our raw reconstruction loss. When compared to sRGB fidelity-only compression, our combined loss leads to appreciable improvements in PSNR of the raw reconstruction with only minor impact on sRGB fidelity as measured by MS-SSIM.

As-projective-as-possible bias correction for illumination estimation algorithms.

Illumination estimation is the key routine in a camera's onboard auto-white-balance (AWB) function. Illumination estimation algorithms estimate the color of the scene's illumination from an image in the form of an R, G, B vector in the sensor's raw-RGB color space. While learning-based methods have demonstrated impressive performance for illumination estimation, cameras still rely on simple statistical-based algorithms that are less accurate but capable of executing quickly on the camera's hardware. An effective strategy to improve the accuracy of these fast statistical-based algorithms is to apply a post-estimate bias-correction function to transform the estimated R, G, B vector such that it lies closer to the correct solution. Recent work by Finlayson [Interface Focus8, 20180008 (2018)2042-889810.1098/rsfs.2018.0008] showed that a bias-correction function can be formulated as a projective transform because the magnitude of the R, G, B illumination vector does not matter to the AWB procedure. This paper builds on this finding and shows that further improvements can be obtained by using an as-projective-as-possible (APAP) projective transform that locally adapts the projective transform to the input R, G, B vector. We demonstrate the effectiveness of the proposed APAP bias correction on several well-known statistical illumination estimation methods. We also describe a fast lookup method that allows the APAP transform to be performed with only a few lookup operations.

Multi-Image Blind Super-Resolution of 3D Scenes.

We address the problem of estimating the latent high-resolution (HR) image of a 3D scene from a set of non-uniformly motion blurred low-resolution (LR) images captured in the burst mode using a hand-held camera. Existing blind super-resolution (SR) techniques that account for motion blur are restricted to fronto-parallel planar scenes. We initially develop an SR motion blur model to explain the image formation process in 3D scenes. We then use this model to solve for the three unknowns-the camera trajectories, the depth map of the scene, and the latent HR image. We first compute the global HR camera motion corresponding to each LR observation from patches lying on a reference depth layer in the input images. Using the estimated trajectories, we compute the latent HR image and the underlying depth map iteratively using an alternating minimization framework. Experiments on synthetic and real data reveal that our proposed method outperforms the state-of-the-art techniques by a significant margin.

Recognizing blurred, nonfrontal, illumination, and expression variant partially occluded faces.

The focus of this paper is on the problem of recognizing faces across space-varying motion blur, changes in pose, illumination, and expression, as well as partial occlusion, when only a single image per subject is available in the gallery. We show how the blur, incurred due to relative motion between the camera and the subject during exposure, can be estimated from the alpha matte of pixels that straddle the boundary between the face and the background. We also devise a strategy to automatically generate the trimap required for matte estimation. Having computed the motion via the matte of the probe, we account for pose variations by synthesizing from the intensity image of the frontal gallery a face image that matches the pose of the probe. To handle illumination, expression variations, and partial occlusion, we model the probe as a linear combination of nine blurred illumination basis images in the synthesized nonfrontal pose, plus a sparse occlusion. We also advocate a recognition metric that capitalizes on the sparsity of the occluded pixels. The performance of our method is extensively validated on synthetic as well as real face data.

Face recognition across non-uniform motion blur, illumination, and pose.

Existing methods for performing face recognition in the presence of blur are based on the convolution model and cannot handle non-uniform blurring situations that frequently arise from tilts and rotations in hand-held cameras. In this paper, we propose a methodology for face recognition in the presence of space-varying motion blur comprising of arbitrarily-shaped kernels. We model the blurred face as a convex combination of geometrically transformed instances of the focused gallery face, and show that the set of all images obtained by non-uniformly blurring a given image forms a convex set. We first propose a non-uniform blur-robust algorithm by making use of the assumption of a sparse camera trajectory in the camera motion space to build an energy function with l1 -norm constraint on the camera motion. The framework is then extended to handle illumination variations by exploiting the fact that the set of all images obtained from a face image by non-uniform blurring and changing the illumination forms a bi-convex set. Finally, we propose an elegant extension to also account for variations in pose.