Search

In this paper, we compress convolutional neural network (CNN) weights post-training via transform quantization. Previous CNN quantization techniques tend to ignore the joint statistics of weights and activations, producing sub-optimal CNN performance at a given quantization bit-rate, or consider their joint statistics during training only and do not facilitate efficient compression of already trained CNN models. We optimally transform (decorrelate) and quantize the weights post-training using a rate–distortion framework to improve compression at any given quantization bit-rate. Transform quantization unifies quantization and dimensionality reduction (decorrelation) techniques in a single framework to facilitate low bit-rate compression of CNNs and efficient inference in the transform domain. We first introduce a theory of rate and distortion for CNN quantization and pose optimum quantization as a rate–distortion optimization problem. We then show that this problem can be solved using optimal bit-depth allocation following decorrelation by the optimal End-to-end Learned Transform (ELT) we derive in this paper. Experiments demonstrate that transform quantization advances the state of the art in CNN compression in both retrained and non-retrained quantization scenarios. In particular, we find that transform quantization with retraining is able to compress CNN models such as AlexNet, ResNet and DenseNet to very low bit-rates (1–2 bits). [ABSTRACT FROM AUTHOR]

This paper describes a visual optimization strategy for scalable video compression. The challenge scalable coding presents is that truncation of an embedded codestream may induce variable and highly visible distortion. To overcome the deficiencies of visually lossless coding schemes, we propose using an adaptive masking slope to model the perceptual impact of suprathreshold distortion arising from resolution and bit-rate scaling. This allows important scene structures to be better preserved. Following visual masking principles, local sensitivity to distortion is assessed within each frame. To keep the perceptual response uniform against spatiotemporal errors, we mitigate errors compounded by the motion field during temporal synthesis. Visual sensitivity weights are projected into the subband domain along motion trajectories via a process called perceptual mapping. This uses error propagation paths to capture some of the noise-shaping effects attributed to the motion-compensated transform. A key observation is that low contrast regions in the video are generally more susceptible to unmasking of quantization errors. The proposed approach raises the distortion-length slope associated with these critical regions, altering the bitstream embedding order so that visually sensitive sites may be encoded with higher fidelity. Subjective evaluation demonstrates perceptual improvement with respect to bit-rate, spatial and temporal scalability. Index Terms--Adaptive masking slope, perceptual coding, scalable video compression, subjective quality evaluation, visual distortion.

Much work has been performed on optimizing the throughput of the block coding system within JPEG2000. However, the question remains as to whether providing parallel simple block coders provides a cheaper method of increasing throughput than complicated optimized block coders. We present the analysis and results for a system on a chip (SoC) software/hardware codesign platform, for parallel coding in JPEG2000 compression standard. We design both a simple and a high performance, optimized peripheral encoder as a hardware accelerator for the JPEG2000 SoC encoding system. The system is implemented on an Altera NIOS II processor with flexible integrated peripheral. We show that there are optimum numbers of parallel block coders and scheduling granularity per row of codeblocks, and that parallel optimized encoders outperform parallel simple encoders. We also demonstrate that the block coding system becomes work starved rather than memory blocked when many parallel coders are present, indicating a discrete wavelet transform bottleneck. Index Terms--Arithmetic coder, block coder, digital architectures, embedded block coder with optimal truncation (EBCOT), JPEG2000, MQ coder, software/hardware codesign, system on a chip.

Discrete wavelet transform (DWT) and embedded block coder (BC) are two main modules in JPEG2000 compression system. Data transfer between the DWT and BC modules presents challenges due to difference in data format generated by the DWT and data format required by the BC module. In this paper, we investigate data transfer and storage techniques between the DWT and BC modules. We propose an efficient memory organization and data transfer schemes to reduce the data bandwidth. A VLSI architecture for the proposed data transfer system is proposed and synthesized for TSMC 0.18-[micro]m process. Simulation results show that our proposed techniques result in approximately four times less bandwidth for the BC module while requiring an extra hardware cost of only 11%. Index Terms--Discrete wavelet transform (DWT), embedded block with optimal truncation (EBCOT), JPEG2000.

Optimal rate-control is an important feature of JPEG2000 which allows simple truncation of compressed bit stream to achieve best image quality at a given target bit rate. Accurate distortion estimation (DE) with respect to the allowed bit stream truncation points, is essential for the rate-control performance. In this paper, we address the issues involved in accurate DE for the hardware oriented implementation of JPEG2000 encoding systems. We propose a novel hardware efficient DE technique. Rate control based on the proposed technique results in a average peak signal-to-noise ratio degradation of only 0.02 dB with respect to the optimal DE technique used in the software implementations of JPEG2000. This is the best performance reported in comparison to existing techniques. The proposed technique requires only an additional 4096 memory bits per block coder which is 80% less than the memory requirements of optimal technique. Index Terms--Distortion estimation (DE), JPEG2000, rate-control.

The block coder, which is a key module in the JPEG2000 image compression system, presents challenges for realization of a high-throughput, low-hardware-cost VLSI architecture. Though efficient architectures have been proposed for a block coder operating in specific modes, existing generic block coder architectures have low throughput versus hardware cost performance. In this paper, we present a low-cost, high-throughput VLSI architecture for a generic block coder. Con-current symbol processing (CSP) is used to improve throughput of the block coder's submodules, the bit plane coder (BPC) and arithmetic coder (AC). The proposed BPC processes one stripe-column/clock-cycle during every coding pass and generates up to 10 context-data (CxD) pairs/clock-cycle. The proposed AC processes two CxD/clock-cycles. Throughput is then further increased by using column speedup and novel run-mode skipping techniques at the BPC module. Hardware cost for the proposed block coder is reduced by using an optimal two-subbank BPC memory architecture. Additionally, image statistics are used to choose efficient configuration parameters for the VLSI architecture. The proposed block coder is implemented on Altera stratix FPGA and TSMC ASIC 0.18-[micro]m platforms. Implementation results show that our block coder has average throughputs of 16.23 and 73.42 Msamples/s, respectively, on the FPGA and ASIC platforms. The block-coder test chip has 22515 gates and 2.33 [mm.sup2] chip area. In comparison with similar existing architectures, it has the highest throughput versus hardware cost performance. Index Terms--Block coder, concurrent symbol processing (CSP), JPEG2000, memory architecture, VLSI architecture.

JPEG2000 is a recently standardized image compression algorithm. The heart of this algorithm is the coding scheme known as embedded block coding with optimal truncation (EBCOT). This contributes the majority of processing time to the compression algorithm. The EBCOT scheme consists of a bit-plane coder coupled to a MQ arithmetic coder. Recent bit-plane coder architectures are capable of producing symbols at a higher rate than the existing MQ arithmetic coders can absorb. Thus, there is a requirement for a high throughput MQ arithmetic coder. We examine the existing MQ arithmetic coder architectures and develop novel techniques capable of absorbing the high symbol rate from high performance bit-plane coders, as well as providing flexible design choices. Index Terms--Arithmetic coding, image compression, JPEG2000.

A unique content adaptive method for motion-compensated three-dimensional wavelet transformation (MC 3 -D DWT) of video is proposed. The results revealed that the adaptive scalable video compression maintains high compression efficiency as well as improves the quality of reduced spatial and temporal resolution sequences.

An investigation is performed on the implications of the conventional 't+2-D' motion-compensated (MC) 3-D discrete wavelet/subband transform structure for spatial scalability and propose a novel flexible structure for fully scalable video compression. The results show that the 't+2-D' structure necessarily maximizes compression efficiency and different schemes are proposed to minimize the loss of compression efficiency associated with pretemporal spatial decomposition.

The dependency structures for scalable data sources and the erasure model are explored. An unequal erasure protection which is able to take advantage of scalable data with more general dependency structures is adapted to data with tree-structured dependencies.

Random accessibility and efficiency enhancements in highly scalable video and volumetric compression are investigated. The impact that the transform, embedded coding components, and code-block configurations have on the compression efficiency and accessibility of a scalable codestream is assessed.

The process in which the priority encoding transmission (PET) framework might be leveraged to exploit both unequal error protection and limited retransmission for RD-optimized delivery of streaming media is illustrated. PET framework allows for a decoupled optimization algorithm with only modest complexity.

A wavelet-based highly scalable video compression scheme with rate-scalable motion coding is presented. This method involves the construction of quality layers for the coded sample data and a separate set of quality layers for the coded motion parameters.

A new framework for highly scalable video compression is proposed using a Lifting-based Invertible Motion Adaptive Transform (LIMAT). A compact representation for the motion parameters, having motion overhead comparable to that of motion-compensated predictive coders is also described.

Scalability refers to the ability to modify the resolution and/or bit rate associated with an already compressed data source in order to satisfy requirements which could not be foreseen at the time of compression. A number of researchers have already demonstrated the feasibility of efficient scalable image and video compression. The principle focus of this paper is to describe data structures for highly scalable compressed video, which are able to support simple, generic scaling approaches for both constant bit rate and constant distortion scaling criteria. Interactive video material presents particular challenges when the data stream is to be scaled to maintain an approximately constant level of distortion, rather than just a constant bit rate. Special attention is paid, therefore, to the development of generic, robust scaling algorithms for such applications. The data structures and scaling methodologies developed in this paper are particularly appealing for distribution of highly scalable compressed video over heterogeneous media, because they simultaneously support both variable bit rate (VBR) and constant bit rate (CBR) services with a wide range of available service qualities, using only simple, generic mechanisms for scaling. The performance of the proposed scaling methodologies is experimentally investigated using a highly scalable video compression algorithm, which is able to achieve comparable compression performance to that of the inherently nonscalable MPEG-1 compression standard.

A new full color video compression technique is developed to produce a single embedded bit stream, which consists of multiple decoder display formats. The technique is based on 3-D subband coding utilizing camera pan compensation. The display formats produce a finely gradated group of bit rates. Single bit streams, which generate suitable subsets, are produced when the technique is experimentally implemented.

A new subband coding technique uses local image orientation to eliminate Gibbs-like phenomena and obtain images with smooth edges even at low bit rates. Index permutation and one-dimensional filtering techniques use the new technique efficiently. The subband-algorithm eliminates all the partitioning artifacts in the images and uses the directional properties of the images.