Real Time Pipelined System for Forming the Sum of Products in the Processing of Video Data

This invention relates to a convolver, namely a system comprising an arithmetic unit for carrying out the operation defined by the following equation: ##EQU1## This equation is explained and illustrated for the case of k=9, i.e., for a 3-by-3 moving window (kernel) of video data and more specifically to a real time pipelined convolver for forming the sum of products p.sub.i w.sub.i, wherein i is a number from 1 to n.sup.2 and n is a number that defines the size of an n-by-n kernel, p.sub.i are the pixel values, typically 12-bit values of an n-by-n kernel, w.sub.i are the convolver weights that may have positive (w.sub.i >0) and negative (w.sub.i <0) values. The pixel values and weights are represented by absolute value binary numbers and a sign bit.

In the processing of video data, it is necessary to produce the sum of products of fixed weights w.sub.i times the corresponding pixel values p.sub.i of successive rows in an n-by-n kernel, such as a 3-by-3 kernel. Examination of the typical weights involved in low-level vision applications indicates that small positive or negative integers are most common, with the ratio of the smallest to the largest weight being usually less than 20. Consequently, each weight contains six bits consisting of a sign bit and five bits for magnitude from zero to 31. The sign bit expands the range of weight values from -31 to +31.

To prevent the data path from overflowing, it is necessary to scale the output of the convolver. Scaling is accomplished by shifting down the data one or more bits, i.e., dividing by some power of two. This is most easily done by switches selecting the parallel output, i.e., by selecting the output from one or more bit positions to the right. The problem is in the requisite hardware for multiplying the pixel values p.sub.i by positive and negative weights w.sub.i and forming the sum of the products.

In accordance with the present invention, a convolver is comprised of N-1 buffers for storing N-1 rasters (scan lines) of pixel values for an n-by-n kernel. The necessary multiplications and additions are carried out in arithmetic operating units connected in cascade in accordance with the following equation: ##EQU2## where k=n.times.n, p.sub.i are the pixel values, w.sub.i are the convolver weights, and p.sub.i are the one's complement of pixel values which are either unsigned or two's complement binary numbers, and weights are represented by absolute value binary numbers plus sign bits.

The number N-1 of complete scan lines that must be stored in order to cover successive n-by-n kernels, are stored in N-1 series connected buffers, each buffer storing a number n of pixels where n is the total number of pixels in a scan line. The accumulation of successive products p.sub.i w.sub.i is accomplished by an array of m one-bit full adders using the absolute values of the convolver weights, both positive and negative, and the one's complement of the pixel value p.sub.i when it is to be multiplied by a negative convolver weight w.sub.i. The accumulated sum of p.sub.i .vertline.w.sub.i .vertline. and p.sub.i .vertline.w.sub.i .vertline. are corrected for the negative convolver weights by adding to the sum of products p.sub.i .vertline.w.sub.i .vertline.+p.sub.i .vertline.w.sub.i .vertline. the sum of .vertline.w.sub.i .vertline. for all convolver weights less than zero, i.e., for negative convolver weights.

The novel features that are considered characteristic of this invention are set forth with particularity in the appended claims. The invention will best be understood from the following description when read in connection with the accompanying drawings.

FIG. 1 is a schematic diagram of the present invention.

FIG. 2 illustrates an example of th algorithm for forming the sum of products PW+S in the system of FIG. 1.

Referring to the drawings, FIG. 1 is a schematic diagram of a 3-by-3 convolver which produces the sum of products of nine fixed weights w.sub.1 through w.sub.9 times the corresponding pixel values p.sub.1 through p.sub.9 in a 3-by-3 moving window of video data. Examination of the typical weights used in low-level vision algorithms indicates that small positive or negative integers are most commonly used, with the ratio of the smallest to the largets weight being usually less than 20. This means that a six-bit (including sign) weight is adequate, since this can repesent integer values from -31 to +31.

To prevent the 12-bit data path from overflowing (because multiplying a 12-bit pixel by a 5-bit weight produces a result PW+S having 17 significant bits), it is also necessary to scale the output of the convolver in some variable way (since all large positive weights will produce a much larger result than a mix of small positive and negative weights). Scaling is most easily accomplished in hardware by shifting the data one or more bits, i.e., dividing by some power of two. If scaling were not included, adding nine of these 17-bit quantities PW+S together can produce a result having as 21 significant bits. Since only 12 of these bits can be output to maintain a constant data path throughout the pipeline, it would be excessive to compute the result accurately to 21 bits. However, somewhat more than 12 bits must be retained in intermediate stages of the convolver, since it is common to take derivatives of heavily smoothed data, which involves subtracting quantities which are nearly equal. To preserve 12 significant bits of result when subtracting quantities differing only by 10%, or so, requires a 16-bit internal data path. To ensure validity of the least significant bit of the output, an additional bit of low significance is also needed internally to the convolver. Thus, the convolver is provided with a 17-bit parallel data path between pipelined blocks for the results PW+S.

Scaling of the 17-bit result is accomplished in two stages: the input pixel values to the convolver may be shifted down in significance, allowing mor room for carry overflow when large positive weights are used, and the 12-bit output actually taken from the 17-bit convolver data output may be shifted up to allow for cancellation when subtracting nearly equal quantities. As discussed above, adding the products of nine 12-by-5 multiplications can produce a 21-bit result. However, having all nine weights near the maximum value of 31 is very unlikely, since they could all be divided by two and the result scaled to produce nearly identical results. Thus, it is reasonable to assume that the sum of the nine weights can always be kept to somewhat under the maximum 279 (9-by-31). If the sum of the weights is kept under 256 (9% less than 279), overflow into the 21st bit can be avoided. This means that shifting the input pixel values down in significance by up to three bits permits the 17-bit data path to accommodate the most significant bits of the 20-bit result. As a consequence, using the convolver of FIG. 1 calls for a programmable shift of from zero to three bits in the input data using external data input select switches as a scaler A. The shift of the output 12-bit data path with respect to the internal 17-bit data path is similarly programmable from zero to three using a scaler B, so that when subtracting nearly equal quantities more significant bits are preserved.

The arithmetic units blocks 10, are preferably implemented with custom VLSI circuit chips. The two complete scan lines that must be stored in order to cover the 3-by-3 window are stored in conventional line buffers (memory chips) comprised of N-3 pixel delays 11 and 12, each with three pixel delay elements in cascade for a total of N pixel delays, and three pixel delay elements which precede the N-3 pixel delay 11 to store the pixels p.sub.1 through p.sub.9 in a moving window array as follows:

p.sub.3 ; p.sub.2 ; p.sub.1

p.sub.6 ; p.sub.5 ; p.sub.4

p.sub.9 ; p.sub.8 ; p.sub.7

It is highly desirable to utilize VLSI technology for implementation of the convolver (except for the line buffers), including registers 13, 14 and 15 for the three sets of weights w.sub.1 -w.sub.3, w.sub.4 -w.sub.6 and w.sub.7 -w.sub.9. Note that the weights are represented by five bits of absolute value (magnitude) plus a sign bit. The sign bits are stored in registers 13a, 14a and 15a which are in reality part of registers 13, 14 and 15, but shown here separate because they are so separated by the convolver algorithm. Custom VLSI implementation allows the 12-by-5 bit arithmetic units 10 to be implemented directly within an internal 17-bit data path. Moreover, a custom VLSI circuit is easiest to design when the circuit to be implemented is a regular, repeated structure, as in this case of pipelined arithmetic units 10.

Accumulation of successive multiplications can be accomplished most straight forwardly by a multiplication algorithm of repeated add and shift operations, as illustrated in FIG. 2 using a simple example of a four-bit pixel as the multiplicand 1010 and a four-bit weight as the multiplier 1101. A nine-bit quantity S equal to 000100110, the sum of absolute values for all negative weights (w.sub.i <0) or the result PW+S from the previous arithmetic unit, is assumed as an input to an arithmetic unit 10. The incoming S is entered in the accumulator of the arithmetic unit as a first step. Then the least significant bit of the multiplier is inspected. Because it is a bit 1, the next step is to add the multiplicand in the four least significant bit positions of the accumulator. In the third step, the next least significant bit of the multiplier is inspected. Because it is a bit 0, zero is effectively added by doing nothing except effectively shifting the multiplicand before going into step number 4. Then the third multiplier bit is inspected. Because it is a bit 1, the multiplicand is effectively shifted, and added to the content of the accumulator. The process is repeated in step 5 for the last bit of the multiplier which, upon inspection after a shift, causes the multiplicand to be added to the partial products in the accumulator. The final product PW+S appears in the accumulator after the last step.

This algorithm for the arithmetic units 10 can be implemented for a 17-bit pixel and 5-bit weight (absolute value) with a 17-by-5 array of one-bit full adder circuits. (Note that in the preferred implementation illustrated in the simple example with reference to FIG. 2, the multiplicand is shifted only in the sense of it being added in a position of one binary bit higher significance in each step using selecting gates to the inputs of the adder, rather than actually shifting the multiplier in the registers 13, 14 and 15.) Thus, nine 17-by-5 arrays, for a total of 765 full adders, are needed for the convolver. In the preferred embodiment, seventeen-bit latches are used at the output of each of the arithmetic units 10 to store the intermediate accumulated results PW+S. This allows each multiplication to take the full pixel scan time. Appropriate reduction in the line buffer delay N corrects for these delays. The multiplication steps needed for each pixel are synchronously performed during pixel scan time.

Signed arithmetic is accomplished by the simple expedient of complementing the pixel value via exclusive OR operators G.sub.1 through G.sub.9 (each consisting of a bank of 17 exclusive OR gates connected to a one pixel delay element D for a 3.times.3 array of data) prior to multiplications by a negative weight. Since the one's complement of the pixel value plus one would produce the negative of that value in two's complement representation, the product of a negative weight and the pixel value is equal to the product of the absolute value of the weight (five bits without a sign bit), and the complement of the pixel value plus one. This is accomplished in the system of FIG. 1 by adding the absolute values of all negative weights together at the outset (since they are fixed at the time of programming), and storing that sum ##EQU3## in a register 16.

The convolver thus organized performs the function of the following equation: ##EQU4## Using two's complement arithmetic, (i.e., -p.sub.i =p.sub.i +1) this function becomes: ##EQU5## However, instead of forming p.sub.i +1, the distributive law is applied yielding the following equation: ##EQU6## where p.sub.i are the pixel values, w.sub.i are the convolver weights, and p.sub.i are the one's complement of the pixel values. In this way a uniform VLSI architecture can handle both positive and negative convolver weights. Note that the input pixel data to the convolver may be positive or negative, and that the final output of the convolver may also be positive or negative. In a particular unit, all bits of a pixel value are complemented when it is to be multiplied by a negative weight expressed in the form of absolute binary value plus a sign bit. Consequently, the multiplicand will have all 1's left of the most significant bit to fill the 17-bit word for the scaled pixel weight (if it was initially positive, or was negative with a positive weight), and so the product PW will be negative for a negative weight, but the output PW+S may be positive or negative depending on the value of S and whether it is positive or negative. In that regard, it should be noted that the correction value ##EQU7## added initially is always a positive number because it is the sum of the absolute value of all negative weights (five bits absolute without the sign bit).

Although particular embodiments of the invention have been described and illustrated herein, it is recognized that modifications and variations may readily occur to those skilled in the art. Consequently, it is intended that the claims be interpreted to cover such modifications and variations.