We propose the structure and color based learned image codec (SLIC) in which the task of compression is split into that of luminance and chrominance. The deep learning model is built with a novel multi-scale architecture for Y and UV channels in the encoder, where the features from various stages are combined to obtain the latent representation. An autoregressive context model is employed for backward adaptation and a hyperprior block for forward adaptation. Various experiments are carried out to study and analyze the performance of the proposed model, and to compare it with other image codecs. We also illustrate the advantages of our method through the visualization of channel impulse responses, latent channels and various ablation studies. The model achieves Bjøntegaard delta bitrate gains of 7.5% and 4.66% in terms of MS-SSIM and CIEDE2000 metrics with respect to other state-of-the-art reference codecs.

JPEG XS is a new standard for low-latency and low-complexity coding designed by the JPEG committee. Unlike former developments, optimal rate distortion performance is only a secondary goal; the focus of JPEG~XS is to enable cost-efficient, easy to parallelize implementations suitable for FPGAs or GPUs. In this article, we shed some light on the entropy coding back-end of JPEG~XS and introduce modifications of the entropy coding stage currently under discussion that improve objective and subjective quality of the compressed images without compromising the parallelism of the original algorithm.

ISO recently published a new image compression standard, JPEG XT, which extends the popular JPEG standard towards higher dynamic range, compression of alpha channels and lossless coding. In part 7 ofJPEG XT, a two-layer lossy image compression for HDR images isintroduced that reconstructs HDR signals by the combination of a baselayer following the legacy JPEG standard, and an extension layer thatenlarges the dynamic range to that of the original image. While bothbase and extension layer are entropy coded by the mechanisms specifiedin the legacy JPEG standard, only the base layer is visible to legacyJPEG decoders. The original JPEG standard now enforces in its part 2error bounds conforming implementations shall satisfy. The questionanswered in this article is in how far these error bounds carry overto error bounds of JPEG~XT for coding of HDR image signals.

The issue of backwards compatible image and video coding gained some attention in both MPEG and JPEG, let it be as extension for HEVC, let it be as the JPEG XT standardization initiative of the SC29WG1 committee. The coding systems work all on the principle of a base layer operating in the low-dynamic range regime, using a tone-mapped version of the HDR material as input, and an extension layer invisible to legacy applications. The extension layer allows implementations conforming to the full standard to reconstruct the original image in the high-dynamic range regime. What is also common to all approaches is the rate-allocation problem: How can one split the rate between base and extension layer to ensure optimal coding? In this work, an explicit answer is derived for a simplified model of a two-layer compression system in the high bit-rate approximation. For a HDR to LDR tone mapping that approximates the well-known sRGB non-linearity of ? = 2.4 and a Laplacian probability density function, explicit results in the form of the Lambert-W-function are derived. The theoretical results are then verified in experiments using a JPEG XT demo implementation.

The issue of backwards compatible image and video coding gained some attention in both MPEG and JPEG, let it be as extension for HEVC, let it be as the JPEG XT standardization initiative of the SC29WG1 committee. The coding systems work all on the principle of a base layer, perating in the low-dynamic range regime, using a one-mapped version of the HDR material as input, and an extension layer invisible to legacy applications. The extension layer allows implementations conforming to the full standard to reconstruct the original image in the high-dynamic range regime. What is also common to all approaches is the rate-allocation problem: How can one split the rate between base and extension layer to ensure optimal coding? In this work, an explicit answer is derived for a simplified model of a two-layer compression system in the high bit-rate approximation. For a HDR to LDR tone mapping that approximates the well-known sRGB non-linearity of gamma = 2.4 and a Laplacian probability density function, explicit results in the form of the Lambert-W-function are derived. The theoretical results are then verified in experiments using a JPEG XT demo implementation.

In its Paris meeting, the JPEG committee decided to work on a backwards compatible extension of the popular JPEG (10918-1) standard enabling lossy and lossless coding of high-dynamic range (HDR) images, the new standard shall allow legacy applications to decompress new code streams into a tone mapped version of the HDR image while codecs aware of the extensions will decompress the stream with full dynamic range. This paper proposes a set of extensions that have rather low implementation complexity, and use - whenever possible - functional design blocks already present in 10918-1. It is seen that, despite its simplicity, the proposed extension performs close to JPEG 2000 (15444-2) and JPEG XR (29199-2) on the HDR test image set of the JPEG for high bit-rates.

As the resolution of monitors and TVs continue to increase, the available bandwidth between host system and monitor becomes more and more a bottleneck. The Video Electronics Standards Association (VESA) is currently developing standards for screen resolutions beyond 4K and, facing the problem of not having enough bandwidth available on traditional copper wires, contacted the JPEG to develop a low complexity, high-throughput still image coder for lossy transmission of video signals. This article describes two approaches to address this predicament, a simple SPIHT based coded and a Hadamard based embedded codec requiring only minimal buffering at encoder and decoder side, and avoiding any pixel-based feedback loops limiting the operating frequency of hardware implementations. Analyzing the details of both implementations reveals an interesting connection between run-length coding, as found in the progressive mode of traditional JPEG coding, and SPIHT/EZW coding - a technique popular in wavelet based compression techniques.

While the context-based entropy coding and bias cancellation steps in the JPEG-LS standard are key features to its compression performance, these steps also enlarge the memory footprint, create dependencies in the data path of implementations and hence limit parallelism in modern multi-core or GPU architectures, and the throughput in hardware implementations. In the proposed modification of JPEG-LS, such most expensive parts with respect to memory space requirements and computational complexity are omitted.

JPEG 2000 part 9, or short JPIP, is an interactive image browsing protocol that allows the selective delivery of image regions, components or scales from JPEG 2000 image. Typical applications are browsing tools for medical databases where transmitting huge images from server to client in total would be uneconomical. Instead, JPIP allows extracting only the desired image parts for analysis by an http type request syntax. Such a JPIP connection may either operate in a session within which the server remains aware of the image data already cached at the client and it hence doesn't have to transmit again, or it may operate in a stateless mode in which the server has no model of the data already available on the client. In such cases, the client may include a description of its cache model within a proceeding request to avoid retransmission of data already buffered. Unfortunately, the standard defined methods how such cache models are described are very inefficient, and a single request including a cache model may grow several KBytes large for typical images and requests, making the deployment of a JPIP server on top of existing http server infrastructure rather inconvenient. In this work, a lossy and loss less embedded compression scheme for such JPIP cache model adjustment requests based on a modified zero-tree algorithm is proposed, this algorithm works even in constraint environments where request size must remain limited. The proposed algorithm losslessly compresses such cache model adjustment requests often better than by a factor of 1:8, but may even perform a 1:8000 compression in cases where the cache model has to describe a large number of precincts.

The JPEG XR image compression solely controls the image quality loss and hence the output rate by means of the quantizer bucket sizes; a precise rate control mechanism like the EBCOT rate allocation algorithm in JPEG 2000 is not specified, and hence rate-distortion optimality of the quantizer is, in general, not given. A simple rate-control mechanism for JPEG XR is introduced that allows an efficient control of the quantizer towards rate-distortion optimality. It was seen in an earlier work that the additional side information required for the spatial varying quantization mechanism of the standard almost compensates the PSNR gain and complicates the rate allocation process by requiring an additional quantizer allocation step.

The JPEG XR image compression standard, originally developed under the name HD-Photo by Microsoft, offers the feature of spatial variably quantization; its codestream syntax allows to select one out of a limited set of possible quantizers per macro block and per frequency band. In this paper, an algorithm is presented that finds the rate-distortion optimal set of quantizers, and the optimal quantizer choice for each macro block. Even though it seems plausible that this feature may provide a huge improvement for images whose statistics is non-stationary, e.g. compound images, it is demonstrated that the PSNR improvement is not larger than 0.3 dB for a two-step heuristics of feasible complexity, but improvements of up to 0.8 dB for compound images are possible by a much more complex optimization strategy.

In this work, we present a SSIM optimal JPEG 2000 rate allocation algorithm. However, our aim is less improving the visual performance of JPEG 2000, but more the study of the performance of the SSIM full reference metric by means beyond correlation measurements.Full reference image quality metrics assign a quality index to a pair of a reference and distorted image. The performance of a metric is then measured by the degree of correlation between the scores obtained from the metric and those from subjective tests. It is the aim of a rate allocation algorithm to minimize the distortion created by a lossy image compression scheme under a rate constraint.Noting this relation between objective function and performance evaluation allows us now to define an alternative approach to evaluate the usefulness of a candidate metric: we want to judge the quality of a metric by its ability to define an objective function for rate control purposes, and evaluate images compressed in this scheme subjectively. It turns out that deficiencies of image quality metrics become much easier visible - even in the literal sense - than under traditional correlation experiments.Our candidate metric in this work is the SSIM index proposed by Sheik and Bovik which is both simple enough to be implemented efficiently in rate control algorithms, but yet correlates better to visual quality than MSE; our candidate compression scheme is the highly flexible JPEG 2000 standard.

Rate allocation in the JPEG2000 image compression algorithm is performed by the EBCOT algorithm, measures file size and distortion, defined as mean square error (MSE). Since MSE correlates only mediocre to visual quality, more advanced metrics like the M-SSIM have been proposed. One exploitable effect of the human visual system is that of visual masking: If a structure of a fixed amplitude is overlayed by a texture, it becomes masked and less visible. This can be addressed in JPEG2000 by multiplying the MSE contribution of a codeblock by a factor mu computed from the neighbourhood of the data. Most of these techniques require, however, complex operations on the coefficients.

Summary form only given. Objective quality assessment of lossy image compression codecs have become an important part of the recent call of the JPEG committee for advanced image coding. We evaluated JPEG with Huffman and arithmetic coding option, a visual and PSNR optimal JPEG2000 version, H.264/AVC and the recently proposed HDPhoto format by Microsoft. For objective evaluation, we use a color version of the M-SSIM metric and the high-dynamic range version of VDP. The results obtained from these tests are compared to subjective testing obtained from an ordering test run by 15 observers in two sessions. Subjective results are compiled to Mean Opinion Score (MOS) and passed through a Kurtosis test to verify their validity and to reject outliers.