Whats a good explanation of the benefits and drawbacks to lossy compression algorithms?

5.4 Summary

Compression algorithms are normally used to reduce the size of a file without removing information. This can increase their entropy and make the files appear more random because all of the possible bytes become more common. The compression algorithms can also be useful when they're used to produce mimicry by running the compression functions in reverse. This is described in Chapter 6.

The Disguise Compression algorithms generally produce data that looks more random. That is, there is a more even distribution of the data.

How Secure Is It? Not secure at all. Most compression algorithms transmit the table or dictionary at the beginning of the file. This may not be necessary because both parties could agree on such a table in advance. Although I don't know how to figure out the mapping between the letters and the bits in the Huffman algorithm, I don't believe it would be hard to figure out.

How to Use It Many compression programs available for all computers. They often use proprietary algorithms that are better than the versions offered here and make an ideal first pass for any encryption program.

Summary

Dictionary compression algorithms use no statistical models. They focus on the memory on the strings already seen. The memory may be an explicit dictionary that can be extended infinitely, or an implicit limited dictionary as sliding windows. Each seen string is stored into a dictionary with an index. The indices of all the seen strings are used as codewords. The compression and decompression algorithm maintains individually its own dictionary but the two dictionaries are identical. Many variations are based on three representative families, namely LZ77, LZ78 and LZW. Implementation issues include the choice of the size of the buffers, the dictionary and indices.

1.1.2 Decompression

Any compression algorithm will not work unless a means of decompression is also provided due to the nature of data compression. When compression algorithms are discussed in general, the word compression alone actually implies the context of both compression and decompression.

In this book, we sometimes do not even discuss the decompression algorithms when the decompression process is obvious or can be easily derived from the compression process. However, as a reader, you should always make sure that you know the decompression solutions as well as the ones for compression.

In many practical cases, the efficiency of the decompression algorithm is of more concern than that of the compression algorithm. For example, movies, photos, and audio data are often compressed once by the artist and then the same version of the compressed files is decompressed many times by millions of viewers or listeners.

Alternatively, the efficiency of the compression algorithm is sometimes more important. For example, the recording audio or video data from some real-time programs may need to be recorded directly to a limited computer storage, or transmitted to a remote destination through a narrow signal channel.

Depending on specific problems, we sometimes consider compression and decompression as two separate synchronous or asynchronous processes.

Figure 1.1 shows a platform based on the relationship between compression and decompression algorithms.

A compression algorithm is often called compressor and the decompression algorithm is called decompressor.

The compressor and decompressor can be located at two ends of a communication channel, at the source and at the destination respectively. In this case, the compressor at the source is often called the coder and the decompressor at the destination of the message is called the decoder. Figure 1.2 shows a platform based on the relationship between a coder and decoder connected by a transmission channel.

There is no substantial difference between the platform in Figure 1.1 and that in Figure 1.2 in terms of the compression algorithms discussed in this book. However, certain concepts may be discussed and understood more conveniently at one platform than the other. For example, it might be easier to introduce the information theory in Chapter 2 based on the coder-decoder platform. Then again, it might be more convenient to discuss the symmetric properties of a compression algorithm and decompression algorithm based on the compressor-decompressor platform.

Compression

Compression algorithms reduce the number of bytes required to represent data and the amount of memory required to store images. Compression allows a larger number of images to be stored on a given medium and increases the amount of data that can be sent over the internet. It relies on two main strategies: redundancy reduction and irrelevancy reduction.

Redundancy reduction, used during lossless encoding, searches for patterns that can be expressed more efficiently. An image viewed after lossless compression will appear identical to the way it was before being compressed. Lossless compression techniques can reduce the size of images by up to half. The resulting compressed file may still be large and unsuitable for network dissemination. However, lossless compression does provide for more efficient storage when it is imperative that all the information stored in an image should be preserved for future use.

Irrelevancy reduction, a lossy compression, utilizes a means for averaging or discarding the least significant information, based on an understanding of visual perception, to create smaller file sizes. Lossy compression reduces the image’s quality but can achieve dramatic storage savings. It should not be used when image quality and integrity are important, such as in archival copies of digital images.

Not all images respond to lossy compression in the same manner. As an image is compressed, particular kinds of visual characteristics, such as subtle tonal variations, may produce what are known as artifacts (unintended visual effects), though these may go largely unnoticed, due to the continuously variable nature of photographic images. Other kinds of images, such as pages of text or line illustrations, will show the artifacts of lossy compression more clearly. These may accumulate over generations, especially if different compression schemes are used, so artifacts that were imperceptible in one generation may become ruinous over many. This is why uncompressed archival master files should be maintained from which compressed derivative files can be generated for access and other purposes.

1.1.3 Measures of Performance

A compression algorithm can be evaluated in a number of different ways. We could measure the relative complexity of the algorithm, the memory required to implement the algorithm, how fast the algorithm performs on a given machine, the amount of compression, and how closely the reconstruction resembles the original. In this book we will mainly be concerned with the last two criteria. Let us take each one in turn.

A very logical way of measuring how well a compression algorithm compresses a given set of data is to look at the ratio of the number of bits required to represent the data before compression to the number of bits required to represent the data after compression. This ratio is called the compression ratio. Suppose storing an image made up of a square array of 256×256 pixels requires 65,536 bytes. The image is compressed and the compressed version requires 16,384 bytes. We would say that the compression ratio is 4:1. We can also represent the compression ratio by expressing the reduction in the amount of data required as a percentage of the size of the original data. In this particular example, the compression ratio calculated in this manner would be 75%.

Another way of reporting compression performance is to provide the average number of bits required to represent a single sample. This is generally referred to as the rate. For example, in the case of the compressed image described above, the average number of bits per pixel in the compressed representation is 2. Thus, we would say that the rate is 2 bits per pixel.

In lossy compression, the reconstruction differs from the original data. Therefore, in order to determine the efficiency of a compression algorithm, we have to have some way of quantifying the difference. The difference between the original and the reconstruction is often called the distortion. (We will describe several measures of distortion in Chapter 8.) Lossy techniques are generally used for the compression of data that originate as analog signals, such as speech and video. In compression of speech and video, the final arbiter of quality is human. Because human responses are difficult to model mathematically, many approximate measures of distortion are used to determine the quality of the reconstructed waveforms. We will discuss this topic in more detail in Chapter 8.

Other terms that are also used when talking about differences between the reconstruction and the original are fidelity and quality. When we say that the fidelity or quality of a reconstruction is high, we mean that the difference between the reconstruction and the original is small. Whether this difference is a mathematical difference or a perceptual difference should be evident from the context.

Abstract

Video compression algorithms rely on spatio-temporal prediction combined with variable-length entropy encoding to achieve high compression ratios but, as a consequence, they produce an encoded bitstream that is inherently sensitive to channel errors. This becomes a major problem when video information is transmitted over unreliable networks, since any errors introduced into the bitstream during transmission will rapidly propagate to other regions in the image sequence.

In order to promote reliable delivery over lossy channels, it is usual to invoke various error detection and correction methods. This chapter firstly introduces the requirements for an effective error-resilient video encoding system and then goes on to explain how errors arise and how they propagate spatially and temporally. We then examine a range of techniques in Sections 11.4, 11.5, and 11.6 that can be employed to mitigate the effects of errors and error propagation. We initially consider methods that rely on the manipulation of network parameters or the exploitation of network features to achieve this; we then go on to consider methods where the bitstream generated by the codec is made inherently robust to errors (Section 11.7). We present decoder-only methods that conceal rather than correct bitstream errors in Section 11.8. These deliver improved subjective quality without adding transmission overhead. Finally in Section 11.9, we describe congestion management techniques, in particular HTTP adaptive streaming (HAS), that are widely employed to support reliable streaming of video under dynamic network conditions.

5.1.3 Implementing other algorithms

The BDI compression algorithm is naturally amenable toward implementation using assist warps because of its data-parallel nature and simplicity. The CABA framework can also be used to realize other algorithms. The challenge in implementing algorithms like FPC [59] and C-Pack [17], which have variable-length compressed words, is primarily in the placement of compressed words within the compressed cache lines. In BDI, the compressed words are in fixed locations within the cache line, and for each encoding, all the compressed words are of the same size and can, therefore, be processed in parallel. In contrast, C-Pack may employ multiple dictionary values as opposed to just one base in BDI. In order to realize algorithms with variable length words and dictionary values with assist warps, we leverage the coalescing/address generation logic [60, 61] already available in the GPU cores. We make two minor modifications to these algorithms [17, 59] to adapt them for use with CABA. First, similar to prior works [17, 54, 59], we observe that few encodings are sufficient to capture almost all the data redundancy. In addition, the impact of any loss in compressibility because of fewer encodings is minimal as the benefits of bandwidth compression are at multiples of a only single DRAM burst (e.g., 32B for GDDR5 [62]). We exploit this to reduce the number of supported encodings. Second, we place all the metadata containing the compression encoding at the head of the cache line to be able to determine how to decompress the entire line upfront. In the case of C-Pack, we place the dictionary entries after the metadata.

We note that it can be challenging to implement complex algorithms efficiently with the simple computational logic available in GPU cores. Fortunately, SFUs [63, 64] are already in the GPU SMs, used to perform efficient computations of elementary mathematical functions. SFUs could potentially be extended to implement primitives that enable the fast iterative comparisons performed frequently in some compression algorithms. This would enable more efficient execution of the described algorithms, as well as implementation of more complex compression algorithms, using CABA. We leave the exploration of an SFU-based approach to future work.

We now present a detailed overview of mapping the FPC and C-Pack algorithms into assist warps.

Advantages and Disadvantages

Each JPEG image is a new fresh image. This is very useful where the frame rate must be very low, such as on an offshore oil platform with a very low-bandwidth satellite uplink, or where only a dial-up modem connection is available for network connectivity. I used JPEG on an offshore platform with only a 64 kb/s satellite connection available. MPEG is most useful where there is adequate data bandwidth available for a fast-moving image but where it is desirable to conserve network resources for future growth and for network stability.

More on Compression Algorithms:

MJPEG: MJPEG is a compression scheme that uses the JPEG-compression method on each individual frame of video. Thus each frame is an entire picture.

MPEG-4: MPEG-4 is a compression scheme that uses a JPEG “Initial Frame (I-Frame),” followed by “Partial Frames (P-Frames),” each of which only addresses the pixels where changes have occurred from the previous frame. After some period of seconds, enough changes have occurred that a new I-frame is sent and the process is started all over again.

H.264: H.264 is a compression scheme that operates much like MPEG-4 but that results in a much more efficient method of storing video, but H.264 relies on a more robust video-rendering engine in the workstation to view the compressed video.

2.2 Moving Picture Experts Group (MPEG) Compression

Standardization of compression algorithms for video was first initiated by CCITT for teleconferencing and video telephony. The digital storage media for the purpose of this standard include digital audio tape (DAT), CD-ROM, writeable optical disks, magnetic tapes, and magnetic disks, as well as communications channels for local and wide area networks, LANs and WANs, respectively. Unlike still image compression, full motion image compression has time and sequence constraints. The compression level is described in terms of a compression rate for a specific resolution.

The MPEG standards consist of a number of different standards. The original MPEG standard did not take into account the requirements of high-definition television (HDTV). The MPEG-2 standards, released at the end of 1993, include HDTV requirements in addition to other enhancements. The MPEG-2 suite of standards consists of standards for MPEG-2 audio, MPEG-2 video, and MPEG-2 systems. It is also defined at different levels to accommodate different rates and resolutions as described in Table 2.

TABLE 2. MPEG-2 resolutions, rates, and metrics [2]

Moving pictures consist of sequences of video pictures or frames that are played back at a fixed number of frames per second. Motion compensation is the basis for most compression algorithms for video. In general, motion compensation assumes that the current picture (or frame) is a revision of a previous picture (or frame). Subsequent frames may differ slightly as a result of moving objects or a moving camera, or both. Motion compensation attempts to account for this movement. To make the process of comparison more efficient, a frame is not encoded as a whole. Rather, it is split into blocks, and the blocks are encoded and then compared. Motion compensation is a central part of MPEG-2 (as well as MPEG-4) standards. It is the most demanding of the computational algorithms of a video encoder.

The established standards for image and video compression developed by JPEG and MPEG have been in existence, in one form or another, for over a decade. When first introduced, both processes were implemented via codec engines that were entirely in software and very slow in execution on the computers of that era. Dedicated hardware engines have been developed and real-time video compression of standard television transmission is now an everyday process, albeit with hardware costs that range from $10,000 to $100,000, depending on the resolution of the video frame. JPEG compression of fixed or still images can be accomplished with current generation PCs. Both JPEG and MPEG standards are in general usage within the multimedia image compression world. However, it seems that the DCT is reaching the end of its performance potential since much higher compression capability is needed by most of the users in multimedia applications. The image compression standards are in the process of turning away from DCT toward wavelet compression.

19.14.3 Compression Algorithms for Packet Video

Almost any compression algorithm can be modified to perform in the ATM environment, but some approaches seem more suited to this environment. We briefly present two approaches (see the original papers for more details).

One compression scheme that functions in an inherently layered manner is subband coding. In subband coding, the lower-frequency bands can be used to provide the basic reconstruction, with the higher-frequency bands providing the enhancement. As an example, consider the compression scheme proposed for packet video by Karlsson and Vetterli [303]. In their scheme, the video is divided into 11 bands. First, the video signal is divided into two temporal bands. Each band is then split into four spatial bands. The low-low band of the temporal low-frequency band is then split into four spatial bands. A graphical representation of this splitting is shown in Fig. 19.31. The subband denoted 1 in the figure contains the basic information about the video sequence. Therefore, it is transmitted with the highest priority. If the data in all the other subbands are lost, it will still be possible to reconstruct the video using only the information in this subband. We can also prioritize the output of the other bands, and if the network starts getting congested and we are required to reduce our rate, we can do so by not transmitting the information in the lower-priority subbands. Subband 1 also generates the least variable data rate. This is very helpful when negotiating with the network for the amount of priority traffic.

Given the similarity of the ideas behind progressive transmission and subband coding, it should be possible to use progressive transmission algorithms as a starting point in the design of layered compression schemes for packet video. Chen, Sayood, and Nelson [304] use a DCT-based progressive transmission scheme [305] to develop a compression algorithm for packet video. In their scheme, they first encode the difference between the current frame and the prediction for the current frame using a 16×16 DCT. They only transmit the DC coefficient and the three lowest-order AC coefficients to the receiver. The coded coefficients make up the highest-priority layer.

The reconstructed frame is then subtracted from the original. The sum of squared errors is calculated for each 16×16 block. Blocks with squared error greater than a prescribed threshold are subdivided into four 8×8 blocks, and the coding process is repeated using an 8×8 DCT. The coded coefficients make up the next layer. Because only blocks that fail to meet the threshold test are subdivided, information about which blocks are subdivided is transmitted to the receiver as side information.

The process is repeated with 4×4 blocks, which make up the third layer, and 2×2 blocks, which make up the fourth layer. Although this algorithm is a variable-rate coding scheme, the rate for the first layer is constant. Therefore, the user can negotiate with the network for a fixed amount of high-priority traffic. In order to remove the effect of delayed packets from the prediction, only the reconstruction from the higher-priority layers is used for prediction.

This idea can be used with many different progressive transmission algorithms to make them suitable for use over ATM networks.