Variable-rate learning excels in standard dynamic range (SDR) image compression, but extending it to high dynamic range (HDR) images is challenging. We propose an end-to-end Variable-Rate Learned HDR (VRLHDR) compression framework.

This paper proposes extensions of CALIC for lossless compression of light field (LF) images. The overall prediction process is improved by exploiting the linear structure of Epipolar Plane Images (EPI) in a slope based prediction scheme. The prediction is improved further by averaging predictions made using horizontal and verticals EPIs. Besides this, the difference in these predictions is included in the error energy function, and the texture context is redefined to improve the overall compression ratio. The results using the proposed method shows significant bitrate-savings in comparison to standard lossless coding schemes and offers significant reduction in computational complexity in comparison to the state-of-the-art compression schemes.

This study presents a framework to predict, in a No Reference (NR) manner, Full Reference (FR) objective quality metrics. The methods are applied to infrared (IR) images acquired by Unmanned Aerial Vehicle (UAV) and compressed on-board and then streamed to a ground computer. The proposed method computes two kinds of features, namely Bitstream Based (BB) features which are estimated from the H.264 bitstream and Pixel Based (PB) features which are estimated from the decoded images. Two BB features are computed using the H.264 Quantization Parameter (QP) and estimated PSNR [1]. A total of 53 PB features are calculated based on spatial information and the rest of the features are based on NR quality assessment methods [1, 2, 3]. The most relevant ones are selected and nally mapped to predict FR objective scores using Support Vector Regression. For the performance evaluation, the proposed method is trained to predict scores of 6 FR image quality metrics (SSIM, NQM, MSSIM, FSIM, MAD and PSNR-HMA) using a set of 250 IR aerial images compressed at 4 levels with H.264/AVC as I-frames. For the SVR mapping, 80% of the contents are used for training (200 contents or 800 images) and the remaining 200 images (20%) for testing. We have evaluated our model for three cases; all features, only BB features and finally excluding BB features. The average SROCC values obtained are 0.970, 0.962 and 0.943, respectively. The BB only version achieves very close results to that of using all features. Thus the presented NR BB Image Quality Assessment (IQA) method for the considered IR image material is very ecient. We have compared our method with three NR methods [1, 2, 3]. The proposed method is competitive compared to the state-of-the-art NR algorithms.

We consider a decomposition method for compressive streaming data in the context of online compressive Robust Principle Component Analysis (RPCA). The proposed decomposition solves an n-ℓ1 cluster-weighted minimization to decompose a sequence of frames (or vectors), into sparse and low-rank components from compressive measurements. Our method processes a data vector of the stream per time instance from a small number of measurements in contrast to conventional batch RPCA, which needs to access full data. The n-ℓ1 cluster-weighted minimization leverages the sparse components along with their correlations with multiple previously-recovered sparse vectors. Moreover, the proposed minimization can exploit the structures of sparse components via clustering and re-weighting iteratively. The method outperforms the existing methods for both numerical data and actual video data.

We consider the task of reconstructing target signals which are processed as sparse sources for a distributed compression scenario, where communication between the sources is prohibited, however, correlation of information among sources can be utilized at the decoder. We propose an efficient reconstruction algorithm with the aid of other given sources as multiple side information (SI) for such distributed sparse sources. The proposed algorithm takes advantage of both a compressive sensing (CS) reconstruction with SI and an iteratively weighted ℓ1-norm minimization by solving a general weighted multi-ℓ1(or n-ℓ1) minimization. To utilize the known multiple SIs, the algorithm computes optimal weights on not only each individual SI but among SIs where the weights are adaptively updated according to changes at every iteration of the reconstruction. By this optimization, the proposed reconstruction algorithm with multiple SI (RAMSI) can robustly exploit the multiple SIs with different qualities. We experimentally demonstrate our algorithm on compressing feature histograms as sparse sources which are extracted from a multi-view image database for multi-view recognition. The results show that the RAMSI with multiple SIs efficiently outperforms the ℓ1minimization and also the CS reconstruction with only one SI.

This paper considers using BCH codes for distributed source coding using feedback. The focus is on coding using short block lengths for a binary source, X, having a high correlation between each symbol to be coded and a side information, Y, such that the marginal probability of each symbol, Xi in X, given Y is highly skewed. In the analysis, noiseless feedback and noiseless communication are assumed. A rate-adaptive BCH code is presented and applied to distributed source coding. Simulation results for a fixed error probability show that rate-adaptive BCH achieves better performance than LDPCA (Low-Density Parity-Check Accumulate) codes for high correlation between source symbols and the side information.

Modern mobile telecommunication systems, such as 3GPP LTE, make use of Hybrid Automatic Repeat reQuest (HARQ) for efficient and reliable communication between base stationsand mobile terminals. To this purpose, marginal posterior probabilities of the received bits are stored in the form of log-likelihood ratios (LLR) in order to combine information sent across different transmissions due to requests. To mitigate the effects of ever-increasing data rates that call for larger HARQ memory, vector quantization (VQ) is investigated as a technique for temporary compression of LLRs on the terminal. A capacity analysis leads to using maximum mutual information (MMI) as optimality criterion and in turn Kullback-Leibler (KL) divergence as distortion measure. Simulations run based on an LTE-like system have proven that VQ can be implemented in a computationally simple way at low rates of 2-3 bits per LLR value without compromising the system throughput.

A new lossless compression scheme for bilevel images targeted at binary shapes of image and video objects is presented. The scheme is based on a local analysis of the digital straightness of the causal part of the object boundary, which is used in the context definition for arithmetic encoding. Tested on individual images of binary shapes and binary layers of digital maps the algorithm outperforms PWC, JBIG and MPEG-4 CAE. On the binary shapes the code lengths are reduced by 21%, 25%, and 42%, respectively. On the maps the reductions are 34%, 32%, and 59%, respectively. The algorithm is also more efficient than the state-of-the-art and more complex free tree coder for most of the binary shape and map test images.

A 2D version of PPM (Prediction by Partial Matching) coding is introduced simply by combining a 2D template with the standard PPM coding scheme. A simple scheme for resolution reduction is given and the 2D PPM scheme extended to resolution progressive coding by placing pixels in a lower resolution image layer. The resolution is increased by a factor of 2 in each step. The 2D PPM coding is applied to palette images and street maps. The sequential results are comparable to PWC. The PPM results are a little better for the palette images with few colors (up to 4-5 bpp) and a little worse for the images with more colors. For street maps the 2D PPM is slightly better. The PPM based resolution progressive coding provides a better result than coding the resolution layers as individual images. Compared to GIF the resolution progressive 2D PPM's coding efficiency is significantly better. An example of combined content-layer/spatial progressive coding is also given.

Context based entropy coding often faces the conflict of a desire for large templates and the problem of context dilution. We consider the problem of finding the quantizer Q that quantizes the K-dimensional causal context C/sub i/=(X(i-t/sub 1/), X(i-t/sub 2/), ..., X(i-t/sub K/)) of a source symbol X/sub i/ into one of M conditioning states. A solution giving the minimum adaptive code length for a given data set is presented (when the cost of the context quantizer is neglected). The resulting context quantizers can be used for sequential coding of the sequence X/sub 0/, X/sub 1/, X/sub 2/, .... A coding scheme based on binary decomposition and context quantization for coding the binary decisions is presented and applied to digital maps and /spl alpha/-plane sequences. The optimal context quantization is also used to evaluate existing heuristic context quantizations.

A new lossless context based method is presented for content progressive coding of limited bits/pixel images, such as maps, company logos, etc., common on the WWW. Progressive encoding is achieved by separating the image into content layers based on other predefined information. Information from already coded layers are used when coding subsequent layers. This approach is combined with efficient template based context bi-level coding, context collapsing methods for multi-level images and arithmetic coding. Relative pixel patterns are used to collapse contexts. The number of contexts are analyzed. The new methods outperform existing coding schemes coding digital maps and in addition provide progressive coding. Compared to the state-of-the-art PWC coder, the compressed size is reduced to 60-70% on our layered test images.

Summary form only given. Lossless image coding may be performed by applying arithmetic coding sequentially to probabilities conditioned on the past data. Therefore the model is very important. A new image model is applied to image coding. The model is based on a Markov process involving hidden states. An underlying Markov process called the slice process specifies D rows with the width of the image. Each new row of the image coincides with row N of an instance of the slice process. The N-1 previous rows are read from the causal part of the image and the last D-N rows are hidden. This gives a description of the current row conditioned on the N-1 previous rows. From the slice process we may decompose the description into a sequence of conditional probabilities, involving a combination of a forward and a backward pass. In effect the causal part of the last N rows of the image becomes the context. The forward pass obtained directly from the slice process starts from the left for each row with D-N hidden rows. The backward pass starting from the right additionally has the current row as hidden. The backward pass may be described as a completion of the forward pass. It plays the role of normalizing the possible completions of the forward pass for each pixel. The hidden states may effectively be represented in a trellis structure as in an HMM. For the slice process we use a state of D rows and V-1 columns, thus involving V columns in each transition. The new model was applied to a bi-level image (SO9 of the JBIG test set) in a two-part coding scheme.

Summary form only given. We investigate lossless coding of video using predictive coding and motion compensation. The new coding methods combine state-of-the-art lossless techniques as JPEG (context based prediction and bias cancellation, Golomb coding), with high resolution motion field estimation, 3D predictors, prediction using one or multiple (k) previous images, predictor dependent error modelling, and selection of motion field by code length. We treat the problem of precision of the motion field as one of choosing among a number of predictors. This way, we can incorporate 3D-predictors and intra-frame predictors as well. As proposed by Ribas-Corbera (see PhD thesis, University of Michigan, 1996), we use bi-linear interpolation in order to achieve sub-pixel precision of the motion field. Using more reference images is another way of achieving higher accuracy of the match. The motion information is coded with the same algorithm as is used for the data. For slow pan or slow zoom sequences, coding methods that use multiple previous images perform up to 20% better than motion compensation using a single previous image and up to 40% better than coding that does not utilize motion compensation.

Presently, tree coders are the best bi-level image coders. The current ISO standard, JBIG, is a good example. By organising code length calculations properly a vast number of possible models (trees) can be investigated within reasonable time prior to generating code. Three general-purpose coders are constructed by this principle. A multi-pass free tree coding scheme produces superior compression results for all test images. A multi-pass fast free template coding scheme produces much better results than JBIG for difficult images, such as halftonings. Rissanen's algorithm 'Context' is presented in a new version that without sacrificing speed brings it close to the multi-pass coders in compression performance.

Partially hidden Markov models (PHMM) are introduced. They are a variation of the hidden Markov models (HMM) combining the power of explicit conditioning on past observations and the power of using hidden states. (P)HMM may be combined with arithmetic coding for lossless data compression. A general 2-part coding scheme for given model order but unknown parameters based on PHMM is presented. A forward-backward reestimation of parameters with a redefined backward variable is given for these models and used for estimating the unknown parameters. Proof of convergence of this reestimation is given. The PHMM structure and the conditions of the convergence proof allows for application of the PHMM to image coding. Relations between the PHMM and hidden Markov models (HMM) are treated. Results of coding bi-level images with the PHMM coding scheme is given. The results indicate that the PHMM can adapt to instationarities in the images.