Image transformations in the compressed domain

[0002] 1. Field of the Invention

[0003] This invention relates generally to the processing of the compressed domain representation of image data, and more particularly to the manipulation of the compressed domain representation to achieve certain spatial domain processing, such as regular geometric transformations of an image, without having to subject the image data to the full decompression and compression process.

[0004] 2. Description of the Related Art

[0005] A typical high quality digitized color image may use 24 bits per pixel (bpp)—8 bits each for red (R), green (G) and blue (B) in RGB color space or for luminance (Y), chrominance (C_B) and chrominance (C_R) in YC_BC_Rcolor space. To transmit or store such images in the uncompressed state (i.e., in spatial or pixel domain) is simply too costly in terms of time and memory requirements. Thus, applications and devices which store and/or transmit high quality digitized color images, such as digital cameras, typically do so in a compressed format, using one of the currently available compression algorithms.

[0006] The emergence of compression standards such as JPEG (an acronym for “Joint Photographic Experts Group”) has led to many digital imaging systems and applications that create and maintain content only in JPEG compressed format. For instance, in most digital still-imaging cameras (DSCs) such as the Epson PhotoPC 600, Kodak DC-10, etc., pictures captured by the camera are immediately compressed within the camera and stored in the camera's storage system as JPEG files. Often there is a need to manipulate these pictures prior to display. Typical image manipulations might include (a) rotating the picture from portrait to landscape mode and vice-versa, (b) scaling the picture to increase or decrease its size, (c) changing brightness and contrast in the picture, (d) cropping portions of the picture for the purposes of creating a new picture and for compositing operations, (e) adding simple bitmap annotations to a picture, and (f) embedding visible/invisible watermarks in the picture. Due to storage constraints within the digital camera, these image manipulations require the processed output to be in JPEG format.

[0007] The need to do these tasks and the availability of the picture only in the compressed mode has resulted in a great deal of interest in developing image processing techniques that can be applied directly to the compressed domain representation. The motivation for investigating compressed domain processing methods stems from the observations that (a) the volume of data in compressed domain tends to be quite small compared to the spatial domain representation which means that fewer operations per sample may be required for the desired image processing task, and (b) conventional processing pipelines that require the data to be decompressed, followed by the application of the desired image processing function in spatial domain, and then recompressed for transmission or storage efficiency can lead to a loss in image fidelity. Furthermore, such a conventional processing pipeline has very high computation complexity or high latency since the compression task is often more complex than the decompression task. The compressed-domain based processing methodology on the other hand, often leads to reduced computation complexity since it replaces the JPEG decompression and compression tasks by low complexity tasks such as Huffman decoding and Huffman encoding. (See, S. F. Chang and D. G. Messerschmitt, “Manipulation and Compositing of MC-DCT Compressed Video,” IEEE JSAC Special Issue on Intelligent Sigital Processing, vol. 13, no. 1, pp. 1-11, January 1995; N. Merhav and V. Bhaskaran, “A fast algorithm for DCT-domain inverse motion compensation,” Proc. ICASSP'96, pp. IV.2307-2310, Atlanta, May 1996; B. Natarajan and V. Bhaskaran, “A fast approximate algorithm for scaling down digital images in the DCT domain,” IEEE International Conference on Image Processing(ICIP), Washington, D.C., October 1995; and Brian Smith and Larry Rowe, “Algorithms for manipulating compressed images,” IEEE Computer Graphics and Applications, pp. 34-42, September 1993.)

[0008] Therefore, it is an object of the present invention to overcome the aforementioned problems associated with performing manipulations on digital images in spatial domain.

[0009] It is another object of this invention to provide for direct manipulation of the compressed domain representation of an image to bring about a selected spatial domain manipulation without having to go through a complete decompression and compression process.

[0010] It is a further object of this invention to provide a set of algorithms which greatly simplify the compressed data manipulations required to bring about the corresponding image manipulations in the spatial domain.

[0011] It is still another object of this invention to provide a set of algorithms to manipulate image data in the compressed domain that includes a manipulation step which takes one set of compressed data and produces another set of compressed data, such that, if the process was reversed, the original image could be produced without any loss of quality.

[0012] The present invention provides a technique for performing various dihedral symmetry operations on a spatial domain representation of a digital image by manipulating a linear transform domain representation of the digital image. When a digital image is in the form of a compressed bitstream, such as a JPEG file, the technique involves entropy decoding the compressed bitstream to generate linear-transform-based data blocks which define the linear transform domain representation of the digital image. The linear-transform-based data blocks are reordered for the corresponding geometric transformation, a particular linear transform domain operation is applied to the data elements within each block, and the blocks are reassembled. When these blocks of data are decompressed into spatial domain, the resulting digital image will be flipped or rotated with respect to the original image. With respect to the original image, the resulting image may be flipped over either diagonal (main or cross-), flipped over either of its middle axes (vertical or horizontal), or rotated 90°, 180° or 270°. The technique may be applied in an imaging device, such as a digital still-image camera, or in a computer system. In either case, the technique may be implemented using hardware or software.

[0013] Other objects and attainments together with a fuller understanding of the invention will become apparent and appreciated by referring to the following description and claims taken in conjunction with the accompanying drawings.

[0014] In the drawings wherein like reference symbols refer to like parts:

[0015] FIG. 1 is a block diagram depicting the JPEG compression and decompression processing flow.

[0016] FIG. 2 is a block diagram showing the spatial-domain based image processing for a JPEG data set.

[0017] FIG. 3 is a block diagram showing the compressed-domain based image processing for a JPEG data set, in accordance with the invention.

[0018] FIG. 4 is a diagram showing a block rotate-by-90° using diagonal-flip (F_d) and a column-flip (F_y) in accordance with the invention.

[0019] FIG. 5 is a block diagram of a digital still-image camera (DSC) which may be used in connection with the invention.

[0020] FIG. 6 is a block diagram that illustrates the interrelationship between various components that may be used in capturing and viewing digital images, as well as processing such images in accordance with the invention.

[0021] Since the 8×8 discrete cosine transform (DCT) is chosen as the basis function in still-image compression standards such as JPEG (see, G. K. Wallace, “The JPEG Still Picture Compression Standard,” Communications of the ACM, vol. 34, no. 4, April 1991), and since JPEG is the compression method widely used in many consumer digital still-cameras in the market, the following description will focus on some of the compressed-domain processing methods that are appropriate for digital still-camera applications and that operate on the DCT domain representations of the camera's JPEG compressed bitstream. Although 8×8 is the most common DCT block size, these techniques may be extended to other DCT block sizes. Furthermore, the techniques are extensible to any other linear transform based basis functions including discrete sine transform, discrete hadamard transform and also the wavelet transforms.

[0022] We begin by briefly describing the JPEG compression and decompression processes and the basic idea of incorporating image processing functions within the JPEG pipeline. JPEG uses the DCT to transform still-image data from its spatial or pixel domain representation to its compressed or frequency domain representation in which the data can be more efficiently coded. The image manipulation methods developed herein are designed to take advantage of the properties of the DCT.

[0023] The JPEG compression and decompression process, illustrated schematically in FIG. 1, operates on a block-by-block basis, where each block size is 8×8. As shown schematically in FIG. 1, the uncompressed still-image 11 is decomposed into 8×8 blocks of pixels by a raster-to-block converter 12. These blocks are then transformed by the forward 8×8 DCT 13 to produce a corresponding set of 8×8 DCT blocks. The forward 8×8 DCT F(u,v) of a spatial domain 8×8 block of samples f(i,j) is computed as: $\begin{array}{cc}F\left(u,v\right)=\frac{{\alpha }_u{\alpha }_v}{4}\sum _{i=0}^7\sum _{j=0}^7{C}_{i,u}{C}_{j,v}f\left(i,j\right),& \left(1\right)\end{array}$

[0024] where, $\begin{array}{cc}{C}_{x,y}=\cos \frac{\left(2x+1\right)y\pi }{12},\mathrm{and},& \left(2\right)\\ {\alpha }_x=\{\begin{array}{cc}1,& x\ne 0\\ 1/\sqrt{2},& \mathrm{otherwise}.\end{array}& \left(3\right)\end{array}$

[0025] After output from the forward 8×8 DCT 13, each of the 64 DCT coefficients is uniformly quantized in a forward quantizer 14 in conjunction with a 64-element quantization table Q, which can be derived empirically to discard information which is not visually significant. In this compression process, the only loss incurred during the compression comes from the quantization of F(u,v) to $F_Q\left(u,v\right)=\mathrm{RoundOff}\left(\frac{F\left(u,v\right)}{Q\left(u,v\right)}\right),$

[0026] where Q is the 8×8 quantization table.

[0027] After quantization, the DCT data in each block is ordered into a “zigzag” sequence which facilitates entropy coding by placing low frequency coefficients (which are more likely to be non-zero) before the high frequency coefficients (which are more likely to be zero). The data is then Huffman coded in a Huffman encoder 15 to further compact the data and to generate a JPEG compressed bitstream.

[0028] The image may be reconstructed from the compressed bitstream using a symmetrical reverse process. The JPEG decompression process begins by decoding the compressed bitstream in a Huffman decoder 16 to regenerate the 8×8 blocks of DCT coefficients. The coefficients are reordered using an inverse zigzaging procedure and the blocks are then fed through an inverse quantizer 17. In the next step, the 8×8 inverse discrete cosine transform (IDCT) 18 operates on the 8×8 blocks of DCT coefficients to generate a stream of 8×8 blocks of pixels. A block-to-raster converter 19 converts these blocks into the decompressed still-image 21. In the decompression process, the IDCT converts the coefficients F(u,v) back to the pixels f(i,j), exactly: $\begin{array}{cc}f\left(i,j\right)=\sum _{u=0}^7\sum _{v=0}^7\frac{{\alpha }_u{\alpha }_v}{4}{C}_{i,u}{C}_{j,v}f\left(u,v\right).& \left(4\right)\end{array}$

[0029] The decompression process will actually work with the quantized coefficients, F_Q, and obtain only an approximation f_Qof f. $\begin{array}{cc}{f}_Q\left(i,j\right)=\sum _{u=0}^7\sum _{v=0}^7\frac{{\alpha }_u{\alpha }_v}{4}{C}_{i,u}{C}_{j,v}{F}_Q\left(u,v\right)Q\left(u,v\right).& \left(5\right)\end{array}$

[0030] If an image processing function such as rotation or scaling has to be performed on an image that is available only as a JPEG compressed bitstream, this processing could be performed as shown in FIG. 2. First, the JPEG compressed bitstream is decompressed in block 31 back into its spatial domain representation. Spatial domain image processing is then performed in block 32. Afterward, the processed pixel data is recompressed in block 33 to generate a new JPEG compressed bitstream. We refer to this scheme as a spatial-domain approach since the processing is directly applied to the decompressed spatial domain data (pixels).

[0031] The spatial-domain approach has certain advantages. One such advantage is that image processing in spatial domain is a well-understood problem and solutions are widely available for many typical image processing functions. Another advantage is that the processing function is independent of the underlying compression scheme used to represent the data.

[0032] This approach has certain disadvantages as well. The data has to be fully decompressed prior to applying the image processing function. Moreover, the processed data may have to undergo the compression process again. Since JPEG is a lossy compression method, decompression-recompression may result in a loss of image quality. For the portrait to landscape conversion application, the quality of an image would incrementally deteriorate every time its orientation is changed. Another disadvantage is that the complexity of the decompression and compression process is quite high. If, for instance, in FIG. 2, the image processing task is, say, clockwise rotation by 90°, then the operations count for rotating each 8×8 block of input data is as listed in Table 1. In order to get a rough estimate of the operations count, we have assumed that multiplies, adds and data accesses can each be done at one operation per data item.

[0033] The present invention proposes an alternative to spatial-domain based image processing, namely, compressed domain based image processing. The latter is well suited for data that is already available in compressed form, such as a JPEG bitstream. The basic processing flow for a compressed-domain based image processing scheme is as depicted in FIG. 3. For JPEG data, compressed-domain based image processing usually implies DCT-domain processing which is represented by block 42 in FIG. 3.

[0034] As shown in FIG. 3, DCT-domain image processing 42 is preceeded by entropy decoding of the JPEG compressed bitstream in block 41 and followed by entropy coding in block 43. As also shown in FIG. 3, the DCT-domain image processing block 42 is further divided into processing blocks labelled 44-49. As previously noted, decompressing the bitstream yields 8×8 blocks of DCT coefficients 44. The coefficients are reordered using a de-zigzag procedure 45, after which the blocks of DCT coefficients are dequantized 46. In accordance with the invention, the dequantized blocks of DCT coefficients undergo the block-oriented processing 47. Following the processing, the DCT-data based blocks are quantized 48 and reordered into the zigzag sequence 49. However, depending on the image processing task, not all of the blocks 44-49 need to be performed. For instance, for the D₄operations (which can be expressed as sequences of flips about the diagonal and the Y-axis), the de-zigzag 45, de-quantize 46, zigzag 49, and quantize 48 steps can be eliminated. The specifics of compressed domain based D₄operations will be discussed in the next section.

[0035] In general, compressed-domain processing has the following advantages. First, image quality could be preserved since in many instances, dequantize-quantize steps can be avoided. Second, the complexity can be much lower than the spatial-domain counterpart depicted in FIG. 2 since full JPEG decompression and compression tasks are avoided. Specifically for the case of rotation by 90°, a simple implementation of the compressed-domain based processing counterpart of Table 1 has an operations count as shown in Table 2. Note that the total operations count using the compressed-domain approach is nearly five times lower than the spatial-domain based approach of Table 1. (Details of the DCT-domain based rotation method are discussed below.) Another advantage is that, in typical digital imagery, due to high correlation among pixels, the DCT-domain representation tends to be quite sparse (e.g., in a 8×8 DCT block, usually there are around 7-16 nonzero values). This data sparseness property can be exploited by the DCT-domain processing approach to further reduce the overall complexity; this property is not available in a spatial-domain representation. (Special techniques to exploit this property for the D₄operations are described below.)

[0036] Note that, in general, it may not be possible to derive a compressed-domain based processing equivalent of a spatial-domain based image processing function. DCT is a linear transform and hence, compressed-domain based processing can probably be accomplished for linear image processing functions. Nonlinear image processing functions such as median-filtering, warping/morphing are not amenable to compressed-domain based approaches such as the one depicted in FIG. 3.

[0037] In this section, we develop the basic equations governing simple geometric transformation of JPEG compressed data. The operations defined by compositions of flips about the diagonal axes, the Y-axis (i.e., the middle-vertical axis) and the X-axis (i.e., the middle-horizontal axis) form the group of dihedral symmetry of the square, referred to as D₄. These operations are listed and described in Table 3.

[0038] Note that the operations F_dand F_ycan be composed to generate the whole group. (Notational convention: the composition o₁o₂of operations o₁and o₂is the operation resulting from first applying o₂and then o₁.) For example, a simple clockwise rotate-by-90° (R₉₀) can be achieved by applying a diagonal-flip followed by a column-flip as depicted in FIG. 4. Thus, if we can derive the compressed-domain based counterparts of the two operations F_dand F_y, all the rest can be derived from them.

[0039] Let f be the 8×8 pixel block, and F be the corresponding 8×8 DCT block (DCT(f)=F and IDCT(F)=f). For any of the eight D₄operations, o, it is easy to express the relationship between f and of. The goal is to derive the relationship between F and DCT(of) (denoted by oF).

[0040] Consider the operation F_y. In spatial domain, the column-flip output F_yf(i,j) can be expressed as:

[0041] F_yf(i,j)=f(i,7−j).  (6)

[0042] From Eq. 1, F_yF(u,v) is: $\begin{array}{cc}{F}_{y}F\left(u,v\right)=\frac{{\alpha }_u{\alpha }_v}{4}\sum _{i=0}^7\sum _{j=0}^7{C}_{i,u}{C}_{j,v}{F}_{y}f\left(i,j\right)& \left(7\right)\\ =\frac{{\alpha }_u{\alpha }_v}{4}\sum _{i=0}^7\sum _{j=0}^7{C}_{i,u}{C}_{j,v}f\left(i,7-j\right)& \left(8\right)\\ =\frac{{\alpha }_u{\alpha }_v}{4}\sum _{i=0}^7\sum _{k=0}^7{C}_{i,u}{C}_{7-k,v}{F}_{y}f\left(i,k\right),& \left(9\right)\end{array}$

[0043] where, k=7−j.

[0044] Using Eq. 2, C_7−k,ycan be expressed as: $\begin{array}{cc}{C}_{7-k,v}=\cos (\frac{\left(2\left(7-k\right)+1\right)v\pi }{16})& \left(10\right)\\ =\cos \left(v\pi -\frac{\left(2k+1\right)v\pi }{16}\right)& \left(11\right)\\ =\cos \left(v\pi \right)\cos \left(\frac{\left(2k+1\right)v\pi }{16}\right)& \left(12\right)\\ ={\left(-1\right)}^v\cos \left(\frac{\left(2k+1\right)v\pi }{16}\right)& \left(13\right)\\ ={\left(-1\right)}^v{C}_{k,v}.& \left(14\right)\end{array}$

[0045] Using Eq. 14 and the DCT definition in Eq. 1, F_yF(u,v) in Eq. 9 can be rewritten as: $\begin{array}{cc}{F}_{y}F\left(u,v\right)=\frac{{\alpha }_u{\alpha }_v}{4}{\left(-1\right)}^v\sum _{i=0}^7\sum _{k=0}^7{C}_{i,u}{C}_{k,v}f\left(i,k\right)& \left(15\right)\\ ={\left(-1\right)}^{v}F\left(u,v\right).& \left(16\right)\end{array}$

[0046] In spatial-domain, diagonal-flip of an input block f(i,j) is F_df(i,j)=f(j,i). It can be easily seen that the DCT-domain equivalent of this is:

[0047] F_dF(u,v)=F(v,u).  (17)

[0048] Using Eq. 15, Eq. 17, and the relationships given in the third column of Table 3, we can derive the compressed domain counterparts for all the D₄operations. These are listed in Table 4.

[0049] The above relationships were derived without taking quantization into account. By simply using the dequantized coefficients, F_Q(u,v)Q(u,v), in place of F(u,v), it can be easily seen that the quantized coefficients of a block produced by a D₄operation can be directly obtained by applying the corresponding input block processing (column 3 of Table 4) on the quantized coefficients of the original block. The quantization table remains the same (it is transposed for F_d, F_cd, R₉₀, and R₋₉₀). This DCT-domain approach avoids the IDCT, the DCT, as well as dequantization and quantization. It will be shown below that the de-zigzag and zigzag steps can also be avoided.

[0050] In order to perform a D₄operation on an W×H JPEG image consisting of many 8×8 blocks (W and H in any JPEG image are multiples of 8, obtained by padding the original image, if necessary), the compressed-domain based procedure is as follows: (a) reorder the 8×8 DCT blocks for the corresponding geometric transformation, and (b) apply the DCT-domain operation as per Table 4 to elements within each 8×8 DCT block. Note that (a) and (b) are lossless operations in that the quantized DCT coefficients F_Q(u,v) are not manipulated beyond a sign-change; by avoiding the process of dequantization and requantization, no quality loss is incurred regardless of the number of times one or more geometric transformations of Table 4 are applied to the camera JPEG file.

[0051] For simplicity, we first describe the implementation techniques using grayscale images; the simple modifications needed for color images are outlined subsequently. We assume that the width and height of the image are multiples of 8. Moreover, for color images, we will assume that subsampling does not necessitate block padding. All the operations described here can also be applied when these constraints are not satisfied, by first padding the original image with some extra rows/columns.

[0052] Consider a W×H grayscale image, I, available as JPEG data. By applying entropy-decoding to the JPEG data, and undoing the differential coding for the DC terms, we can obtain the quantized DCT coefficients for each block. Let F_kdenote the 8×8 block (numbered k in raster-order) of quantized DCT coefficients for the image (0≦k<WH/64).

[0053] Let I^odenote the result of applying operation o on the image (where o is one of the D₄operations from Table 3). From the preceding discussion, it is apparent that the DCT coefficient blocks of I^owill essentially be the same as those in I, with possible reordering, transposition, and sign changes, and the quantization table will also be the same, with possible transposition. In general, the block F_k^oof quantized coefficients in I^o, can be expressed as:

[0054] F_k^o=oF_p^o(k)

[0055] where p^ois a permutation of the blocks. Consider, as an example, clockwise rotation by 90° (o=R₉₀). For k=i_b(H/8)+j_b(that is the block in row i_band column j_bof blocks, in the rotated image I^o), we have,

[0056] p^o(k)=(H−j_b−1)(W/8)+i_b.

[0057] In order to create the JPEG image I^o, we need to calculate the blocks F_k^oin the order, k=0, 1, 2, . . . This would require accessing the blocks of I in the order p^o(0), p^o(1), p^o(2), . . . , which is (in general) different from the raster-order in which these blocks are stored in the JPEG image I. Since the compressed blocks can be of arbitrarily different sizes, extraction of a particular block from the JPEG bitstream would require parsing of all the prior blocks, which is too expensive. Moreover, the value of the quantized DC coefficient in any block can only be extracted after all the previous DC values are decoded, because of differential coding. We avoid these problems by using a two-pass approach. In the first pass, the JPEG image I is parsed to extract the bit-offset for each block and the DC value for each block. In the second pass, when computing F_o^kfor the image I^o, the block F_p^o(k) is readily accessed by looking up the bit-offset for block number p^o(k) and seeking to that position in the JPEG bitstream for I. This algorithm is summarized in the following pseudo-code.

[0058] procedure OperateJPEG

[0059] input: JPEG image I, operation o

[0060] output: JPEG image I^o

[0061] /* first pass: gather offsets and DC values */parse off the header from I

[0062] for k=0 to WH/64−1

[0063] 1. store the offset into I as the bit-offset for block k

[0064] 2. parse I to move past another block while extracting the DC term

[0065] 3. undo the differential coding and store the DC value for block k

[0066] /* second pass: compute and store I^o*/compute and store header for I^o, transposing the quantization table of I if required by o

[0067] for k=0 to WH/64−1

[0068] 1. use the offset and DC tables to seek into I and extract the quantized coefficients of block number p^o(k)

[0069] 2. transpose and/or apply sign changes to the coefficients, as required by o (Table 4, column 3)

[0070] 3. append the entropy-coded coefficients to the JPEG bitstream for I^o

[0071] The heart of the “OperateJPEG” algorithm lies in the second pass, when the original block F_p^o_(k)is transformed via o to get the block F_k^ofor the image I^o. Transposing and applying sign-changes to an 8×8 block would require accessing each element of the block once, if implemented in a straight-forward manner. That is, the complexity will be a multiple of 64. The quantized coefficient blocks in typical JPEG image have a very small number (typically less than 16) of non-zero coefficients. We now present an algorithm to do the transposition and sign-changes with a complexity proportional to the number of non-zero coefficients in the block. This results in a substantial decrease in the overall complexity, making it linear in the total size of the compressed image, rather than in the total size of the uncompressed image.

[0072] In the Huffman coding mode of JPEG a block of quantized coefficients is coded by scanning it in zigzag order to group together long runs of zeros. The zigzag ordering is specified in Table 5. When coding a block, first the difference between the quantized DC value of the current block and the previous block is coded, using a Huffman table specific for these differentials.

[0073] Next, the quantized AC coefficients are scanned in zigzag order and coded as a sequence of symbol trios of the form (R, S, V), where R is the number of consecutive zeros in the zigzag ordering, and the next non-zero coefficient has the value x such that:

[0074] S=┌log₂(|x|+1)┐,

[0075] $V=S\mathrm{least}\mathrm{significant}\mathrm{bits}\mathrm{of}\{\begin{array}{cc}x& \mathrm{when}x>0\\ x-1& \mathrm{when}x<0.\end{array}$

[0076] The bits for V are extracted using the standard 2's complement binary representation of x or x−1. In the JPEG bitstream, each (R, S, V) is coded using first a Huffman table to encode (R, S) and then S extra bits to encode V There are some special coding situations such as very long runs of zeros and end-of-block, handled by special codes. But for our purposes, it should be clear from this description that the JPEG data can be easily parsed to capture each block in a data structure of the form:

[0077] In this data structure, N is the number of non-zero AC coefficients in the block. Only the first N entries of the array A are meaningful. The element A[k] gives the zig-zag location (Z), and the S and V values for the k^thnon-zero AC coefficient, 0≦k<N. Reading a block from a JPEG bitstream into this structure, and writing this structure as JPEG data, are both straightforward, and the details are omitted here.

[0078] The operation o may require transposition and/or sign changes. We now describe a novel algorithm for implementing o, using the above data structure, such that only N steps are needed (rather than 64). This algorithm avoids de-zigzagging and zigzagging of the coefficients, as it uses the zigzag-ordered block representation directly.

[0079] Given an input “JpegBlock” structure B (corresponding to a block F), we would like to find the structure B^o, corresponding to oF. For changing the sign of a non-zero quantized AC coefficient, x, with the corresponding S and V values being S_xand V_x, it is sufficient simply to take the bitwise complement of V_x. That is,

[0080] S_−x=S_x, V_−x=−V_x.

[0081] Let s^o[64] be a precomputed array of boolean flags such that s^o[Z] is TRUE if and only if o requires the sign of the Z^thzigzag coefficient to be changed. If only sign-changes are needed (i.e., for the operations F_x, F_y, R₁₈₀), then, B can be converted into B^oby copying each entry, while flipping the bits for those A[k]. V for which s^o[A[k].Z] is TRUE.

[0082] Transposing a “JpegBlock” structure in N steps uses the following key observation: the array A remains the same, except that elements on any given cross-diagonal are reversed in order. To illustrate this, let t[Z] denote the transposed zigzag index for the original zigzag index Z. That is,

[0083] t[1]=2, t[2]=1, t[3]=5, t[4]=4, . . .

[0084] Consider a situation when the zigzag coefficients numbered:

[0085] 1, 3, 6, 7, 10, 11, 14

[0086] are the only non-zero coefficients. Then, the zigzag ordering after transposition will be:

[0087] t[1], t[3], t[7], t[6], t[4], t[10].

[0088] Thus, each group of coefficients lying on the same cross-diagonal needs to be reversed. This can be done efficiently by scanning the array B.A[. . . ] of non-zero coefficients, pausing at each coefficient where the cross-diagonal changes to copy the previous cross-diagonal from B.A[ . . . ] to B^o.A[ . . . ] in reverse order.

[0089] The following piece of pseudo-code summarizes the algorithm. The arrays t and s^o(defined previously) are precomputed. In addition, let d[64] be another precomputed array, giving the “cross-diagonal number” for each zigzag index. That is, if the zig-zag index Z corresponds to row number u and column number v, then, d[Z]=u+v.

[0090] procedure TransposeAndSignChange

[0091] input: JpegBlock B, arrays t, s^o, d

[0092] output: JpegBlock B^o

[0093] In general, a JPEG image I consists of more than one color plane, with some planes subsampled. Let the number of color planes be denoted by P. Associated with each plane, p (I≦p≦P), there is a horizontal sampling factor, w_p, and a vertical sampling factor, h_p. The width and height of plane number p are given by W w_p/w_maxand H h_p/h_max, respectively. Here, w_maxis the maximum value of w_p, and h_maxis the maximum value of h_p, over all planes (I≦p≦P).

[0094] The JPEG data can have a further layer of structuring, consisting of scans. A scan consists of one or more color planes, with possibly a limited bit-precision of coefficient values. The “OperateJPEG” algorithm can be extended simply by applying it to each scan in succession.

[0095] Within each scan, the data from the blocks of all the color planes in the scan is organized into units known as minimum coded units (MCUs). Each MCU consists of a fixed number of blocks in a fixed order, determined by all the w_pand h_p. For the operations we are considering, the image I^oconsists of exactly the same scans and MCU groupings as I, as long as we do not change the sampling factors (h_pand w_pdo need to be swapped for the operations requiring transposition). The order of the MCUs and the order of blocks within each MCU might change, but each MCU in I^ocan be obtained from exactly one MCU in I. We exploit this fact to also save some memory by building the bit-offset and DC tables at MCU level, rather than block level.

[0096] Thus, for every scan in I, “OperateJPEG” works in two passes. In the first pass, for each MCU in the scan, the bit-offset is recorded, and the DC value of the first block in that MCU from each color plane is recorded. In the second pass, when a particular MCU for the scan of I^ois to be computed, the corresponding MCU of I is extracted and the “JpegBlock” structure for each constituent block is filled out. These blocks are then transposed/sign-changed/reordered (as required by the operation) to form the MCU for I^owhich is written out as JPEG data.

[0097] The compressed-domain processing techniques of the present invention may be employed in connection with various digital devices including a digital still-image camera (DSC), a block diagram of which is illustrated in FIG. 5. Operating under microprocessor control, the DSC 60 has a charge-coupled device (CCD) image sensor that captures an image and converts it to an analog electrical signal in block 61. The analog signal is then processed and digitized in block 62, after which the digital image is temporarily stored in a frame buffer 63 while it undergoes digital processing in block 64. The digital image processing block 64 performs several functions including compression and decompression and may also perform the compressed-domain based processing techniques of the present invention. Under user control 65, the processing block 64 interfaces with in-camera image storage 66 where decompressed image data may be stored. The storage block 66 may comprise compact magnetic or solid-state storage media, either removable or fixed within the DSC 60, and may include removable, large-capacity PCMCIA-format hard disk cards or flash memory cards.

[0098] The DSC 60 includes analog and digital outputs, 67 and 68 respectively, through which image data may be transmitted within the DSC or to external devices. Uncompressed image data may be transmitted, via the analog outputs 67, to an LCD screen 69 within the DSC 60, or to external devices such as a VCR or TV monitor. Image data, whether compressed or uncompressed, may also be transmitted through the digital outputs 68 to a digital device such as a computer system where the image could be displayed.

[0099] The ability to perform the D₄operations such as rotation by 90°,180°, 270° or mirror-flip by directly manipulating the DCT-domain representations is quite useful in the DSC 60. When a camera is held vertically to capture a tall object, the resulting image appears as an image that has undergone a 90° rotation with respect to the same picture taken with the camera held horizontally. With typical film cameras this is not a problem, since when holding the prints, one simply rotates the picture by 90° to view the correct image. With DSCs 60 pictures taken in the manner described above can be problematic since often these images are directly printed or viewed on a TV monitor or incorporated in a computer document. By incorporating compressed-domain based image processing, during readout of the DSC image file, the DSC 60 can produce the corrected view of the image by undoing the rotation as per the methods described herein. In a DSC system, the rotation function can be within the DSC 60 or within the software driver running on a computer which is connected to the DSC 60. For a DSC 60 having limited computing capabilities and limited memory, a compressed-domain based approach might be the only efficient way in which to perform the geometric transformation. The alternative approach of decompressing the data and then performing the manipulation is memory intensive and the slower CPU speed in the DSC 60 can lead to significant performance penalties. As DSC systems evolve, additional features such as image enhancement and watermarking can be easily incorporated within the compressed domain processing framework without having to go back-and-forth between a compressed domain representation and spatial-domain representation.

[0100] As noted above, the DSC 60 shown in FIG. 5 can also be used in connection with a computer system and other components in capturing, processing and viewing digital images. FIG. 6 is a block diagram that illustrates the interrelationship between the DSC 60, a computer system and various other components. The computer system, identified generally by reference numeral 100, may be of any suitable type such as a main frame or personal computer.

[0101] Computer system 100 comprises a central processing unit (CPU) 101 which may be a conventional microprocessor, a random access memory (RAM) 102 for temporary storage of information, and a read only memory (ROM) 103 for permanent storage of information. Each of these components is coupled to a bus 104. Operation of the computer system 100 is typically controlled and coordinated by operating system software. The operating system, which is embodied in the system memory and runs on CPU 101, coordinates the operation of computer system 100 by controlling allocation of system resources and performing a variety of tasks, such as processing, memory management, networking and I/O functions, among others.

[0102] Also coupled to bus 104 by a controller 105 is a diskette drive 106 into which a non-volatile mass storage device such as a diskette 107 may be inserted. Similarly, a controller 108 interfaces between bus 104 and a compact disc (CD) ROM drive 109 which is adapted to receive a CD ROM 110. A hard disk 111 is provided as part of a fixed disk drive 112 which is coupled to bus 104 by a disk controller 113.

[0103] Software for the compressed-domain based processing techniques may be stored on storage devices 107 and 110 and transferred to CPU 101 for execution. Alternatively, the software may be stored in RAM 102 or ROM 103. Similarly, image data processed or to be processed in accordance with the invention may be loaded into and extracted from computer system 100 using removable storage media devices such as the diskette 107 and CD ROM 110.

[0104] Image data may be input into computer system 100 in other ways as well. Film-based images 114 generated by a film camera 115 can be digitized by a scanner 116 for storage and processing by the computer 100. The DSC 60 can directly digitize images and transmit them to the computer 100, as previously explained. A keyboard 121 and mouse 122, which are coupled to bus 104 via a controller 123, facilitate the input of such data and otherwise provide a means for entering information into computer system 100.

[0105] Image data may also be transferred to and from computer 100 for remote locations. To this end, computer 100 may also include a communications adapter 124 which enables the computer 100 to communicate with networks 125, which may include local area networks (LANs), the internet or online services, via direct connections or via modem.

[0106] Digital images transmitted or stored in computer 100 may be viewed in a number of different ways. A printer 126 attached to computer 100 can produce color prints that vary in quality depending on the printer 126. Another option is to view the images on a display 127 associated with the computer 100. Yet another choice is to display the images on a television receiver using a VCR.

[0107] As the foregoing description demonstrates, the inventors herein have developed a simple compressed-domain based processing framework for JPEG compressed still imagery. It has been shown that simple geometric transformations such as mirror image, rotations by 90°, 180° and 270° can be easily performed in the DCT domain without any loss of image fidelity. Furthermore, it was also shown that even a simple transformation such as rotate-by-90° can achieve significant speedup when performed in DCT domain instead of the conventional spatial-domain based processing approach. The practical use of the compressed-domain based image transformations in a real-world setting, namely a digital still-camera has also been described. Since images captured by a digital camera are quite noisy to begin with, it is imperative that any image processing applied to these images should not increase the noise and the methods developed in this report are essentially lossless operations. Furthermore, these methods are well suited to the limited computing and memory capabilities found in consumer digital still cameras. The basic compressed domain framework can be used for other types of image transformations; specifically, the methodology developed here can be extended to image enhancement, image filtering (see, B. Chitprasert and K. R. Rao, “Discrete Cosine Transform Filtering,” Signal Processing, vol 19, pp. 233-245, 1990), resolution translation, etc.

[0108] While the invention has been described in conjunction with specific embodiments, it will be evident to those skilled in the art in light of the foregoing description that many further alternatives, modifications and variations are possible. For example, the block diagrams used to illustrate the compressed-domain based processing techniques of the present invention, show the performance of certain specified functions and relationships thereof. The boundaries of these functional blocks have been arbitrarily defined herein for the convenience of description. Alternate boundaries may be defined so long as the specified functions and relationships thereof are appropriately formed. Moreover, the psuedo-code used to illustrate the algorithms of the present invention does not depict syntax or any particular programming language. Rather, it provides the functional information one skilled in the art would require to fabricate circuits or to generate software to perform the processing required. Each of the functions depicted in the block diagrams may be implemented, for example, by software instructions, a functionally equivalent circuit such as a digital signal processor circuit, an application specific integrated circuit (ASIC) or combination thereof. The present invention, having been thus described, is intended to embrace all such alternatives, modifications, applications and variations as may fall within the spirit and scope of the appended claims.