In this letter, we propose a novel scheme for the compression of the bilevel region of interest (ROI) maps of hyperspectral images. The scheme can achieve high compression ratios by novel ways of partitioning the intricate regions of the image into smaller blocks. We first analyzed the effect of a low-order linear predictive model on the original image. The analysis showed that linear prediction results in lower-entropy residual images compared to the original images, which in turn helps in improving the compression efficiency. We then use the optimization algorithm that finds the best combination of scan directions for the nonzero blocks. It is shown to offer increasingly better compression with additional iterations. The simulation results on various data sets show that our scheme outperforms the international standard for binary image compression [joint bi-level image expert (JBIG2)].

Thelow-complexity RAHT (Region-Adaptive Hierarchical Transform) and its corresponding Dyadic decomposition have been incorporated in the MPEG (Moving Picture Expert Group) standardization for encoding the attribute values in 3D point clouds. While significant gains were observed with the replacement of RAHT by Dyadic RAHT, the standard scheme uses only one type of decomposition for the entire point cloud. In this letter, we propose a novel adaptive scheme of switching between RAHT and Dyadic RAHT to achieve early termination of Dyadic RAHT by using 3D gradient filters. Improvements were made on the state-of-the-art scheme that uses only Dyadic RAHT transformation. The proposed method was tested on the publicly available 3D point cloud datasets. The results shows we can achieve cumulative compression gain of up to 11% over all-dyadic approach, by avoiding further decomposition in some cases.

While Golomb-Rice codes are optimal for geometrically distributed source, the practically achievable coding efficiency depends on the accuracy of the coding parameter estimated from the input data. Most existing methods are based on the assumption of geometric distribution and thus would suffer from a loss in coding efficiency if the underlying distribution deviates from the geometric distribution, which is usually the case in practice. We proposed a data-driven parameter estimation method without assuming the underlying distribution. We formulated the problem of choosing the best coding parameter for the given input data as a pattern classification problem. To this end, we trained a deep belief network using the data segments to be coded, along with their “labels”, which are the optimal coding parameters that yield the shortest codewords. Simulations on data synthesized using statistical models, as well as data in hyperspectral image coding showed that the proposed deep learning method tended to be more robust than several state-of-the-art parameter estimation methods, with the capability to further improve the accuracies of these methods.

This paper addresses the problem of efficient predictive lossless compression on the regions of interest (ROIs) in the hyperspectral images with no-data regions. We propose a two-stage prediction scheme, where a context-similarity-based weighted average prediction is followed by recursive least square filtering to decorrelate the hyperspectral images for compression. We then propose to apply separate Golomb-Rice codes for coding the prediction residuals of the full-context pixels and boundary pixels, respectively. To study the coding gains of this separate coding scheme, we introduce a mixture geometric model to represent the residuals associated with various combinations of the full-context pixels and boundary pixels. Both information-theoretic analysis and simulations on synthetic data confirm the advantage of the separate coding scheme over the conventional coding method based on a single underlying geometric distribution. We apply the aforementioned prediction and coding methods to four publicly available hyperspectral image data sets, attaining significant improvements over several other state-of-the-art methods, including the shape-adaptive JPEG 2000 method.

We propose a novel predictive lossless compression algorithm for regions of interest (ROIs) in the hyperspectral images via Maximum Correntropy Criterion (MCC) based Least Mean Square (LMS) filtering. Non-linearity and non-Gaussian conditions of prediction residuals of the ROI pixels in the hyper-spectral image are taken into account to improve the compression performance compared to the ordinary LMS used in the Consultative Committee for Space Data Systems (CCSDS) standard. Test results on hyperspectral image datasets show that the proposed method outperforms several other state-of-the-art methods.

While autoencoders have been used as an unsupervised machine learning technique for classification and dimensionality reduction of the input data, they are lossy in nature when used alone in data compression. In this work, we proposed an image coding scheme by using stacked autoencoders, where the reconstruction residuals were entropy-coded to achieve lossless compression. As a case study, we compressed labeled red blood cell images from a database curated by pathologists for malaria infection diagnosis. Specifically, we trained two separate stacked autoencoders to automatically learn the discriminative features from input images of infected and non-infected cells. Subsequently, the residuals of these two classes of images were coded by two independent Golomb-Rice encoders. Testing results showed that this deep learning approach provided remarkably higher compression on average than several other lossless coding methods including JPEG-LS, JPEG 2000 lossless mode, and CALIC.

Block-based motion estimation is widely employed by many video compression systems to capture temporal correlations of the data. While the exhaustive full search block matching method can find globally optimal candidates without any loss in the matching accuracy, high computational complexity of this method limits its practical applications. Lossless block matching algorithms based on successive candidate elimination were found to be effective in reducing the complexity of the full search method. In this paper, we introduce a new breadth-first method with multiple passes to speedup the depth-first multi-level successive elimination algorithm (MSEA), by exploiting the correlations between matching metrics of different levels. Simulations showed that the proposed lossless block-matching method based on MSEA reduced the running times of the conventional full search method by about 17.5 times and 9.5 times on average for QCIF and CIF sequences, respectively, representing about 11% and 34% improvements over the MSEA.

Fast and accurate block-based motion estimation (BME) is desired in many video coding systems. By conducting block matching in lower dimensional projection space, followed by candidate exclusion through thresholding, projection-based BME (PBME) methods can run several times faster than the exhaustive full search method, with little loss in accuracy. In PBME methods, the appropriate choice of threshold is critical, as the threshold controls the important trade-offs between complexity and accuracy. In the literature, PBME methods rely on fixed thresholds that were chosen on an ad hoc basis. This paper provides an in-depth analysis of errors due to thresholding in PBME, based on the sum absolute difference (SAD) as the matching criterion. A new PBME method is proposed that can adaptively select thresholds according to a target probability of error. Simulation results show that this adaptive method can offer highly scalable complexity/accuracy tradeoffs desired in many fast BME algorithms.

We propose a novel computation control method for variable-complexity Fano decoders with buffers. Our method substantially lowers the rates of data block loss associated with conventional Fano decoders. For reasonably large buffer sizes, our method outperforms Layland's buffer management scheme with block loss rates close to the theoretical lower bound.

The minimization process of the Levenberg-Marquardt algorithm (LMA) used in estimating the global motion parameters tends to be very expensive computationally due to the involvement of all the pixels of an image frame. We propose to reduce the computational complexity of the LMA by using only a small portion of the image data in two stages. In the first stage, we seek to reduce the complexity of the initial guess of the transformation parameters, which is critical to the final convergence of the algorithm. The complexity of computing the initial guess can be lowered by using just a small subset of the pixels in the calculation of the translational components. The second stage of the LMA algorithm is to find the final motion parameters in an iterative fashion, based on the coarse estimate of the motion parameters obtained in the previous stage. The LMA in this stage again operates on a subset of the pixels to further reduce the overall computational complexity. Both analytical and simulation results showed that the proposed partial-data algorithm could achieve a speedup factor of over 25 for global motion estimation (GME) with an eight-parameter perspective motion model on several video sequences, without significant loss in the estimation accuracy compared with the conventional LMA on the full image data.

Global motion generally describes the motion of the camera, although it may comprise large object motion. The region of support for global motion representation consists of the entire image frame. Therefore, estimating global motion parameters tends to be computationally costly due to the involvement of all the pixels in the calculation. Efficient global motion estimation (GME) techniques are sought after in many applications such as video coding, image stabilization and super-resolution. In this paper, we propose to select only a small subset of the pixels in estimating the global motion parameters, based on a combination of fixed and random subsampling patterns. Simulation results demonstrate that the proposed method was able to speed up the conventional all-pixel GME approach by up to 7 times, without significant loss in the estimation accuracy. The combined subsampling patterns were also found to provide better motion estimation accuracy/complexity tradeoffs than those achievable by using either fixed or random patterns alone.

In portable mobile communications, variable complexity Fano decoders are of interest because they offer a desirable tradeoff between bit error rate (BER) and decoding complexity. The practical use of Fano decoders is made possible by buffers that are able to absorb the variations in decoding delays. The optimal buffer control scheme aims to minimize the overall probability of block loss, which is comprised of the loss due to excessive bit errors caused by fast yet coarse decoding, and the loss due to buffer overflow caused by long processing delays. While our previous work has yielded a solution to the control problem over memoryless channel, we now investigate Fano decoding over channels with memory. We seek to achieve better control performance by taking advantage of the memory of the channel. We use slow, flat Rayleigh fading channels as an example to demonstrate that prediction based on two-state Markov models can further lower the probability of block loss.

Fano sequential decoders are variable complexity convolutional decoders, which have the desirable property of operating with very low computation at high SNR. In portable mobile communications, it is often desirable to trade BER with decoder complexity/power consumption. However, the variable complexity nature of the Fano algorithm means that buffers are required for the Fano decoder due to large variations in processing delays. In this paper, we formulate the buffer control problem as one that seeks to minimize the overall probability of block loss, subject to a finite buffer size constraint. The overall probability of block loss is comprised of two terms, corresponding to loss due to excessive bit errors and decoder buffer overflow, respectively. This leads to an interesting trade-off, as faster decoding often means higher bit error rate. Based on the joint distribution of decoding complexity and BER, at each decoding stage, we find an optimal Fano decoder parameter (/spl Delta/) to minimize block loss. In addition, we propose a simple real-time table lookup algorithm that implements the /spl Delta/ control policy. Simulation results demonstrate the superior performance of the proposed algorithm.

In applications where compression has to be performed under varying complexity constraints (e.g., with hardware having to operate in reduced power mode) it is beneficial to design compression algorithms that allow some degree of complexity scalability. In this paper we explore complexity scalability for transform coding algorithms. We show that a variable complexity algorithm (VCA), which uses energy thresholds to determine the number of coefficients to be computed for each input, is preferable to other alternatives such as a pruned transform, where the same number of coefficients is computed for the whole image. We show that the benefits include not only a higher degree of scalability, but also increased compression performance, as we take advantage of the energy classification that is needed for VCA operation and design quantizers that match each class. We provide expressions for the average complexity as well as rate/distortion relations for a generic N-point VCA transform. For a two point case, we present closed-form relations describing the variance changes in two classes. In addition, rate-distortion-complexity relations are also empirically obtained. We apply VCA to eight-point KLT and 8/spl times/8 DCT in the JPEG framework and experiments show that the VCA approach is superior in rate/distortion performance at low rates compared to the standard transform coding techniques.

The complexity of the discrete cosine transform (DCT) is a concern in portable video compression devices, where multiplications are typically much more costly than additions and binary shifts. Since most conventional fast DCT algorithms exploit the algebraic structure of the DCT, their multiplicative complexity has been shown to have lower bounds. In this paper, we take advantage of the distributed arithmetic (DA) structure of the 2-D DCT. We introduce a novel fast DA-DCT algorithm based on DA optimization, which reduces the number of additions by a factor of 22, through recursive pairwise matching. On average, only 1 multiplication, 40 additions as well as 16 binary shifts are required for each DCT coefficient. The overall multiplicative complexity is 28% lower than the theoretical lower bound. Our DA-DCT is numerically equivalent to the exact, double precision floating-point 2-D DCT.

A high performance 1D-DCT is proposed. It is based on a new distributed arithmetic architecture technique (NEDA). Only addition operations are used, with 35 additions to complete the first phase of the computations. The final phase is the primitive add-and-shift operation. No subtraction, multiplication, or ROM are needed. High-throughput is achieved by pipelining the architecture. The delay of one stage is the delay of one 12-bit addition. Low power consumption is achieved by reducing the computation requirements compared to other implementations of the 1D-DCT.