Tài liệu Cơ sở dữ liệu hình ảnh P9 pdf

In addition to image volume both image size and number of images, other factors are competing for transmission bandwidth, including the continued growth in demand for access and the comp

Edited by Vittorio Castelli, Lawrence D Bergman Copyright  2002 John Wiley & Sons, Inc ISBNs: 0-471-32116-8 (Hardback); 0-471-22463-4 (Electronic)

JEFFREY W PERCIVAL

University of Wisconsin, Madison, Wisconsin

VITTORIO CASTELLI

IBM T.J Watson Research Center, Yorktown Heights, New York

The transmission of digital imagery can impose significant demands on the resources of clients, servers, and the networks that connect them With the explo-sive growth of the internet, the need for mechanisms that support rapid browsing

of on-line imagery has become critical

This is particularly evident in scientific imagery The increase in availability of publicly accessible scientific data archives on the Internet has changed the typical scientist’s expectations about access to data In the past, scientific procedures produced proprietary (PI-owned) data sets that, if exchanged at all, were usually exchanged through magnetic tape Now, the approach is to generate large on-line data archives using standardized data formats and allow direct access by researchers For example, NASA’s Planetary Data System is a network of archive sites serving data from a number of solar system missions The Earth-Observing System will use eight Distributed Active Archive Centers to provide access to the very large volume of data to be acquired

Although transmission bandwidth has been increasing with faster modems for personal networking, upgrades to the Internet, and the introduction of new networks such as Internet 2 and the Next Generation Internet, there are still

a number of reasons because of which bandwidth for image-transmission will continue to be a limited resource into the foreseeable future

Perhaps the most striking of these factors is the rapid increase in the resolution of digital imagery and hence the size of images to be transmitted A case in point is the rapid increase in both pixel count and intensity resolution provided by the solid-state detectors in modern astronomical systems For example, upgrading from a

640× 480 8-bit detector to a 2,048 × 2,048 16-bit detector represents a data growth

of about a factor of 30 Now, even 2,0482detectors seem small; new astronomical

241

detectors are using 2,048 × 4,096 detectors in mosaics to build units as large as 8,1922 pixels This is a factor of 400 larger than the previously mentioned 8-bit

image The Sloan Digital Sky Survey will use thirty 2,0482detectors simultaneously and will produce a 40 terabyte data set during its lifetime

This phenomenon is not limited to astronomy For example, medical imaging detectors are reaching the spatial and intensity resolution of photographic film and are replacing it Similarly, Earth-observing satellites with high-resolution

sensors managed by private companies produce images having 12,0002 pixels, which are commercialized for civilian use

In addition to image volume (both image size and number of images), other factors are competing for transmission bandwidth, including the continued growth

in demand for access and the competition for bandwidth from other application such as telephony and videoconferencing Therefore, it does not seem unrea-sonable to expect that the transmission of digital imagery will continue to be a challenge requiring careful thought and a deft allocation of resources

The transmission of digital imagery is usually handled through the exchange of raw, uncompressed files, losslessly compressed files, or files compressed with some degree of lossiness chosen in advance at the server The files are first transmitted and then some visualization program is invoked on the received file Another type of transmission, growing in popularity with archives of large digital images, is called

progressive transmission When an image is progressively transmitted from server

to client, it can be displayed by the client as the data arrive, instead of having to wait until the transmission is complete This allows browsing in an archive even over connections for which the transmission time of a single image may be prohibitive

In this chapter each of these transmission schemes and their effect on the allocation

of resources between server, network, and client are discussed

This is the simplest case to consider Error-free transmission of raw digital images

is easily done using the file transfer protocol (FTP) on any Internet-style (TCP/IP) connection Image compression can be used to mitigate the transmission time by decreasing the total number of bytes to be transmitted

When used to decrease the volume of transmitted data, the compression usually needs to be lossless, as further data analysis is often performed on the received data sets A lossless compression is exactly reversible, in that the exact value

of each pixel can be recovered by reversing the compression Many compres-sion algorithms are lossy, that is, the original pixel values cannot be exactly recovered from the compressed data Joint photographic experts group (JPEG)[1] and graphics interchange format (GIF) (see Chapter 8) are examples of such algorithms Lossy compression is universally used for photographic images trans-mitted across the World Wide Web Interestingly, it is becoming increasingly important in transmitting scientific data, especially in applications in which the images are manually interpreted, rather than processed, and where the bandwidth between server and client is limited

Compression exploits redundancy in an image, which can be large for certain kinds of graphics such as line drawings, vector graphics, and computer-generated images Lossless compression of raw digital imagery is far less efficient because digital images from solid-state detectors contain electronic noise, temperature-dependent dark counts, fixed-pattern noise, and other artifacts These effects reduce the redundancy, for example, by disrupting long runs of pixels that would otherwise have the same value in the absence of noise A rule of thumb is that lossless compression can reduce the size of a digital image by a factor of 2 or 3 The cost of transmitting compressed images has three components: the cost of compression, the cost of decompression, and the cost of transmission The latter decreases with the effectiveness of the compressed algorithm: the bigger the achieved compression ratio, the smaller the transmission cost However, the first two costs increase with the achieved compression ratio: when comparing two image-compression algorithms, one usually finds that the one that compresses better complex is more1 Additionally, the computational costs of compressing and decompressing are quite often similar (although asymmetric schemes exist where compression is much more expensive than decompression) In an image database, compression is usually performed once (when the image is ingested into the database) and therefore its cost is divided over all the transmissions of the image Hence, the actual trade-off is between bandwidth and decompression cost, which depends on the client characteristics Therefore, the compression algorithm should be selected with client capabilities in mind

A number of high-performance, lossless image-compression algorithms exist Most use wavelet transforms of one kind or another, although simply using a wavelet basis is no guarantee that an algorithm is exactly reversible A well-tested, fast, exactly reversible, wavelet-based compression program is HCOMPRESS[2]

Source code is available at www.stsci.edu.

Finally, it is easily forgotten in this highly networked world that sometimes more primitive methods of bulk transmission of large data sets still hold some sway For example, the effective bandwidth of a shoebox full of Exabyte tapes sent by overnight express easily exceeds 100 megabits per second for 24 hours

Progressive transmission is a scheme in which the image is transmitted from server to client in such a way that the client can display the image as it arrives, instead of waiting until all the data have been received

Progressive transmission fills an important niche between the extremes of transmitting either raw images in their entirety or irreversibly reduced graphic products

1 This assertion needs to be carefully qualified: to compress well in practice, an algorithm must be tailored to the characteristics of actual images For example, a simple algorithm that treats each pixel

as an independent random variable does not perform as well on actual images as a complex method that accounts for dependencies between neighboring pixels.

Progressive transmission allows users to browse an archive of large digital images, perhaps searching for some particular features It has also numerous applications in scientific fields Meteorologists often scan large image archives looking for certain types or percentages of cloud cover The ability to receive, examine, and reject an image while receiving only the first 1 percent of its data

is very attractive Astronomers are experimenting with progressive techniques for remote observing They want to check an image for target position, filter selection, and telescope focus before beginning a long exposure, but the end-to-end network bandwidth between remote mountain tops and an urban department

of astronomy can be very low Doing a quality check on a 2,0482 pixel, 16-bit image over a dial-up transmission control protocol (TCP) connection seems daunting, but it is easily done with progressive image-transmission techniques Some progressive transmission situations are forced on a system The Galileo probe to Jupiter was originally equipped with a 134,400 bit-per-second transmission system, which would allow an image to be transmitted in about a minute The high-gain antenna failed to deploy; however, it resulted in a Jupiter-to-Earth bandwidth

of only about 10 bits per second Ten days per image was too much! Ground system engineers devised a makeshift image browsing system using spatially subsampled

images (called jail bars, typically one or two image rows spaced every 20 rows) to

select images for future transmission Sending only every twentieth row improves the transmission time, but the obvious risk of missing smaller-scale structure in the images is severe Figure 9.1 shows the discovery image of Dactyl, the small moon orbiting the asteroid Ida Had Dactyl been a little smaller, this form of image browsing might have prevented its discovery

Ideally, progressive transmission should have the following properties

• It should present a rough approximation of the original image very quickly It should improve the approximation rapidly at first and eventually reconstruct the original image

Figure 9.1 Discovery image of Dactyl, a small moon orbiting the asteroid Ida Had the

moon been a little smaller, it could have been missing in the transmitted data.

• It should capture features at all spatial and intensity scales early in the transmission, that is, broad, faint features should be captured in the early stages of transmission as easily as bright, localized features

• It should support interactive transmission, in which the client can use the first approximations to select “regions of interest,” which are then scheduled for transmission at a priority higher than that of the original image By

“bootstrapping” into a particular region of an image based on an early view, the client is effectively boosting bandwidth by discarding unneeded bits

• No bits should be sent twice As resolution improves from coarse to fine, even with multiple overlapping regions of interest having been requested, the server must not squander bandwidth by sending the client information that it already has

• It should allow interruption or cancellation by the client, likely to occur while browsing images

• It should be well behaved numerically, approximately preserving the image statistics (e.g., the pixel-intensity histogram) throughout the transmission This allows numerical analysis of a partially transmitted image

Progressive image transmission is not really about compression Rather, it is better viewed as a scheduling problem, in which one wants to know which bits

to send first and which bits can wait until later Progressive transmission uses the same algorithms as compression, simply because if a compression algorithm can tell a compressor which bits to throw away, it can also be used to sense which bits are important

9.3.1 Theoretical Considerations

In this section, we show that progressive transmission need not require much more bandwidth than nonprogressive schemes

To precisely formulate the problem, we compare a simple nonprogressive scheme to a simple progressive scheme using Figure 9.2 Let the nonprogressive scheme be ideal, in the sense that in order to send an image with distortion (e.g.,

mean-square error, MSE, or Hamming distance) no larger than D N, it needs to

send R N bits per pixel and that the point (R N , D N ) lies on the rate-distortion curve[3] defined in Chapter 8, Section 8.3 No other scheme can send fewer bits and produce an image of the same quality

The progressive scheme has two stages In the first, it produces an image

having distortion no larger than D1 by sending R1 bits per pixel and in the

second it improves the quality of the image to D2 = D N by sending further

R2 bits per pixel Therefore, both schemes produce an image having the same quality

Our wish list for the progressive scheme contains two items: first we would like to produce the best possible image during the initial transmission, namely, we

wish the point (R1, D1) to lie on the rate-distortion curve, namely R1= R(D1)

(constraint nr 1), second, we wish overall to transmit the same number of bits

0 1 2 3 4 5 6 7 8 9 10 0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Distortion D (MSE)

R(D)

(RN,DN) = (R1+R 2 ,DN) Second stage

First stage

Ideal behavior of progressive transmission

(R1,D1)

Figure 9.2 Ideal behavior of a successive refinement system: the points describing the

first and second stage lie on the rate-distortion curve R(D).

R N as the nonprogressive scheme, that is, we wish R1+ R2 = R N = R(D N )=

R(D2)(constraint nr 2)

In this section it is shown that it is not always possible to satisfy both constraints and that recent results show that constraints 1 and 2 can be relaxed

to R1< R(D1) + 1/2 and R1+ R2≤ R(D2) + 1/2, respectively.

The first results are due to Equitz and Cover [4], who showed that the answer depends on the source (namely, on the statistical characteristics of the image), and that although, in general, the two-stage scheme requires a higher rate than the one-stage approach, there exist necessary and sufficient conditions under which

R1= R(D1) and R1+ R2 = R(D2) (the sources for which the two equalities

hold for every D1and D2are called successively refinable2) An interesting ques-tion is whether there are indeed sources that do not satisfy Equitz and Cover’s

2 More specifically, a sufficient condition for a source to be successively refinable is that the original

data I , the better approximation I2, and the coarse approximation I1form a Markov chain in this

order, that is, if I1 is conditionally independent of I given I2 In simpler terms, the conditional independence condition means that if we are given the finer approximation of the original image, our uncertainty about the coarser approximation is the same regardless of whether we are given the original image.

conditions Unfortunately, sources that are not successively refinable do exist:

an example of such a source over a discrete alphabet is described in Ref [3], whereas Ref [5] contains an example of a continuous source that is not succes-sively refinable This result is somewhat problematic: it seems to indicate that the rate-distortion curve can be used to measure the performance of a progressive transmission scheme only for certain types of sources

The question of which rates are achievable was addressed by Rimoldi [6], who

refined the result of Equitz and Cover: by relaxing the condition that R1= R(D1) and (R1+ R2) = R(D2), the author provided conditions under which pairs of rates R1and R2can be achieved, given fixed distortions D1and D2 Interestingly,

in Ref [6], D1 and D2 need not be obtained with the same distortion measure This is practically relevant: progressive transmission methods in which the early stages produce high-quality approximations of small portions of the image and poor-quality renditions of the rest are discussed later (e.g., in telemedicine, a radiographic image could be transmitted with a method that quickly provides

a high-quality image of the area of interest and a blurry version of the rest of the image [7], which can be improved in later stages Here, the first distortion measure could concentrate on the region of interest, whereas subsequent measures could take into account the image as a whole.)

Although Rimoldi’s regions can be used to evaluate the performance of a progressive transmission scheme, a more recent result [8] provides a simpler answer For any source producing independent and identically distributed samples3 under squared error distortion for an m-step progressive transmission scheme, any fixed set of m distortion values D1> D2 > D m, Lastras and

Berger showed that there exists an m-step code that operates within 1/2 bit of the

rate-distortion curve at all of its steps (Fig 9.3) This is a powerful result, which essentially states that the rate-distortion curve can indeed be used to evaluate a progressive transmission scheme: an algorithm that at some step achieves a rate

that is not within 1/2 bit of the rate-distortion curve is by no means optimal and

can be improved upon

The theory of successive refinements plays the same role for progressive trans-mission as the rate-distortion theory for lossy compression

In particular, it confirms that it is possible, in principle, to construct progressive transmission schemes that achieve transmission rates comparable to those of nonprogressive methods

This theory also provides fundamental limits to what is achievable and guide-lines to evaluate how well a specific algorithm performs under given assumptions Such guidelines are sometimes very general and hence difficult to specialize to actual algorithms However, the field is relatively young and very active; some of the more recent results can be applied to specific categories of progressive trans-mission schemes, such as the bounds provided in Ref [9] for multiresolution coding

3 Lastras’s thesis also provides directions on how to extend the result to stationary ergodic sources.

0 1 2 3 4 5 6 7 8 9 10 0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Distortion D (MSE)

R(D) +1/2 R(D)

(R1+R 2 ,DN) Second stage

(R1,D1)

First stage

1/2 bit Attainable behavior of progressive transmission

Figure 9.3 Attainable behavior of a progressive transmission scheme: each stage is

described by a point that lies within 1/2 bits of the rate-distortion curve.

Note, finally, that there is a limitation to the applicability of the theory of successive refinements to image transmission: a good distortion measure that is well-matched to the human visual system is not known The implications are discussed in Section 9.3.3

9.3.2 Taxonomies of Progressive Transmission Algorithms

Over the years, a large number of progressive transmission schemes have appeared in the literature Numerous progressive transmission methods have been developed starting from a compression scheme, selecting parts of the compressed data for transmission at each stage, and devising algorithms to reconstruct images from the information available at the receiver after each transmission stage The following taxonomy, proposed by Tsou [10] and widely used in the field, is well suited to characterize this class of algorithm because it focuses on the characteristics of the compression scheme Tsou’s taxonomy divides progressive transmission approaches into three classes:

Spatial-Domain Techniques Algorithms belonging to this class compress

images without transforming them A simple example consists of dividing the

image into bit planes, compressing the bit planes separately, and scheduling the transmission to send the compressed bit planes in order of significance Dividing

an 8-bit image I into bit planes consists of creating eight images with pixel

values equal to 0 or 1 — the first image contains the most significant bit of the

pixels of I , the last image contains the least significant bit, and the intermediate

images contain the intermediate bits

A more interesting example is progressive vector quantization (VQ): the image

I is first vector-quantized (Chapter 8, Section 8.6.2) at low rate, namely, with

a small codebook, which produces an approximation I1 The difference I − I1

between the original image and the first coarse approximation is further quantized,

possibly with a different codebook, to produce I2, and the process is repeated

The encoded images I1, I2, are transmitted in order and progressively recon-structed at the receiver

Transform-Domain Techniques Algorithms belonging to this category

trans-form the image to a specific space (such as the frequency domain), and compress the transform Examples of this category include progressive JPEG and are discussed in Section 9.4.1

Pyramid-Structured Techniques This category contains methods that rely on

a multiresolution pyramid, which is a sequence of approximations of the

orig-inal image I at progressively coarser resolution and larger scale (i.e., having

smaller size) The coarsest approximation is losslessly compressed and trans-mitted; subsequent steps consist of transmitting only the information necessary

to reconstruct the next finer approximation from the received data Schemes derived from compression algorithms that rely on subband coding or on the wavelet transform (Chapter 8, Sections 8.5.3 and 8.8.2) belong to this category, and in this sense, the current category overlaps the transform-domain techniques Numerous progressive transmission algorithms developed in recent years are not well categorized by Tsou’s taxonomy

Chee [11] recently proposed a different classification system, which uses Tsou’s taxonomy as a secondary categorization Chee’s taxonomy specifically addresses the transmission mechanism, rather than the compression characteristics, and contains four classes:

Multistage Residual Methods This class contains algorithms that progressively

reduce the distortion of the reconstructed image Chee assigns to this class only methods that operate on the full-resolution image at each stage: multiresolution-based algorithms are assigned to the next category This category includes multi-stage VQ [12] and the transform-coding method proposed in Ref [13]

Our discussion of successive refinements in Section 9.3.1 is directly relevant

to this class of methods

Hierarchical Methods These algorithms analyze the images at different scales

to process them in a hierarchical fashion Chee divides this class into nonresidual coder, residual multiscale coders, and filter-bank coders

• Nonresidual coders perform a multiscale decomposition of the image and

include quadtree-coders [14,15], binary-tree coders [16], spatial pyramid coders [17,18], and subsampling pyramids [19]

• Residual coders differ from nonresidual coders in that they compute and

encode the residual image at each level of the decomposition (the difference between the original image and what is received at that stage) The well-known Laplacian pyramid can be used to construct a hierarchical residual coder [20] A theoretical analysis of this category of coders can be found

in Ref [21]

• Filter-bank coders include wavelet-based coders and subband coders.

Wavelet-based coders send the lowest resolution version of the image first and successively transmit the subbands required to produce the approximation at the immediately higher resolution A similar approach is used in subband coders [22] The theoretical analysis of Ref [9] is directly relevant to this group of methods

Successive Approximation Methods This class contains methods that

progres-sively refine the precision (e.g., the number of bits) of the reconstructed approx-imations Methods that transmit bit planes belong to this category: at each stage the precision of the reconstructed image increases by 1 bit Chee assigns to this category bit planes methods, tree-structured vector quantizers, full-search quan-tizers with intermediate codebooks [23], the embedded zerotree wavelet coder (Chapter 8, Section 8.8.2), and the successive approximation mode of the JPEG standard (Section 9.4.1.1)

Note that these methods, and in particular, transform-domain methods, are not guaranteed to monotonically improve the fidelity of the reconstructed image at each stage, if the fidelity is measured with a single-letter distortion measure (i.e.,

a measure that averages the distortions of individual pixels) [1]

Methods Based on Transmission Sequences In this category, Chee groups

methods that use a classifier to divide the data into portions, prioritize the order

in which different portions are transmitted, and include a protocol for specifying transmission order to the receiver The prioritization process can aim at different goals, such as reducing the MSE or improving the visual appearance of the reconstructed image at each step

In Ref [11] the author assigns to this class the spectral selection method of the JPEG standard (Section 9.4.1.1), Efstratiadi’s Filter Bank Coder [24], and several block-based spatial domain coders [25,26]

9.3.3 Comparing Progressive Transmission Schemes

Although it is possible to compare different progressive transmission schemes [11], no general guidelines exist However, the following broad statements can be made: