I NASA Technical Memorandum 102325 Digital CODEC for Real-Time Processing of Broadcast Quality Video Signals at 1 8 Bits/Pixel Mary Jo Shalkhauser and Wayne A. Whyte, Jr. Lewis Research Center Cleveland, Ohio (NASA-Tll-102325) DIGITAL CODEC FOB lV8 9 -279 2 7 REAL-TIHE PBOCESSIIG OF BRO&DCAST QUALITY VIDgO SIGWALS AT 1.8 BITS/PIXBL IMASa- Lewis Research Center) 16 p CSCL 17B Dnclas G3/32 0225953 Prepared for the * Global Telecommunications Conference sponsored by the Institute of Electrical and Electronics Engineers Dallas, Texas, November 27-30, 1989 i .. ,_, DIGITAL CODEC FOR REAL-TIME PROCESSING OF BROADCAST QUALITY VIDEO SIGNALS AT 1.8 BITSIPIXEL Mary Jo Shalkhauser and Wayne A. Whyte. Jr. National Aeronaut cs and Space Administration Lewi s Research Center C1 eve and, Ohio 44135 11. Abstract signals in the same bandwidth occupied by a single frequency modulated television signal. This paper Advances in very large-scale integration presents the hardware implementation of a digital and recent work in the field of bandwidth television bandwidth compression algori thm which efficient digital modulation techniques have processes standard NTSC (National Television Systems combined to make digital video processing Committee) composite color television signals and technically feasible and potentially cost com- produces broadcast quality video in real time at an petitive for broadcast quality television average of 1.8 bits/pixel. (A pixel, or picture transmission. A hardware implementation has element, represents each piece of sampled data. been developed for a DPCM-based digital televi- The sampling rate used with this algorithm results sion bandwidth compression algorithm which in 768 samples over the active portion of each processes standard NTSC composite color video line by 512 active video lines per video television signals and produces broadcast frame.) The algorithm is based on differential quality video in real time at an average of pulse code modulation (DPCM), but additionally 1.8 bitslpixel. This paper describes the data utilizes a nonadaptive predictor, nonuniform quan- compression algorithm and the hardware imple- tizer and multilevel Huffman coder to reduce the mentation of the codec, and provides perform- data rate substantially below that achievable with ance results. straight DPCM. The nonadaptive predictor and multi- level Huffman coder combine to set this technique apart from prior-art DPCM encoding algorithms. 11. Introduction Section I11 below will provide the details of the compression algorithm while sections IV and V dis- Transmission of television signals in a digital cuss the hardware implementation and performance format has been looked upon with promise for a results, respectively. number of years. Digital systems providing tele- conferencing quality video have become common place 111. Data Compression Algorithm in both government and industry. However, digital transmission of high-quality (toll grade or broad- Differential pulse code modulation has histor- cast quality) television signals has yet to achieve ically been one of the most popular predictive anything close to the same kind of acceptance. image coding methods studied, due to its simplicity This has been due in part to the broadcasters' of implementation and overall subjective perform- reluctance to have processing of any kind performed ance characteristics. The fault of DPCM schemes in on the transmitted signals. But to a greater the past have been that 3 to 4 bitslpixel were extent, digital transmission of broadcast quality required to achieve acceptable image quality, with video has failed to gain acceptance because it has 4 bits/pixel generally preferred to maintain a not been cost effective to do so. The lack of broadcast qual i ty picture representation. The sys- available wideband digital links as well as the tem presented here combines the simplicity of the complexity of implementation of bandwidth effi- basic DPCM approach with several performance cient digital video CODEC's (encoder/decoder) has enhancements to achieve broadcast quality images worked to keep the cost of digital television at an average 1.8 bits/pixel. transmission too high to compete with analog , methods. A block diagram of the compression scheme is presented in Fig. 1. The DPCM portion utilizes an Advances in vsiry large-scale integration intrafield approach with a two-dimensional predic- (VLSI) as well as recent work in the field of tio? based on averaging neighboring pixel values advanced digi tal modul$.rlon techniques have com- having the same color subcarrier phase relationship bined to make dtqqtal video processing technically as the current pixel . Sampling of the compos1 te feasible and potqntially cost competitive for analog video signal is done at four times the color broadcast quality television transmission. The subcarrier frequency rate (4 x 3.579545 MHz). Fig- coupling 'of a transparent, bandwidth efficient data ure 2 shows the spacial relationship of the current compression technique with a bandwidth efficient pixel and the two pixels used to generate the pre- modulation technique offer the potential for trans- dicted value, PV, in Fig. 1. The pixels used are mission of two (or more) high-quality television the fourth previous pixel from the same line and 1 I I I I I DIGITIZED VIDEO IN I QUANTIZER I VALUE I I I I I I I I NON-ADAPTIVE I I ~~~~ RP I PREDICTOR I I I I o;p I I I I I LINE LINE I +2 I + I DELAY DELAY I ENCODER I I I I I NON-ADAPTIVE I I PREDICTOR I I I I I I I RECONSTRUCTED I DlGlTlZEDVlDEO 4 1 OUT I I I I I I I I I I I I I I I I I I I I DECODER I SAMPLES r2LINES PREVIOUS PIXEL CONSECUTIVE LINESOFA -0 0 0 0 0 0 0 0 FIELD -o~oooqoo 4TH CURRENT PIXEL PREVIOUS PIXEL Figure 2. - Anabg video signal sampling and DPCM pixel relationships. the same pixel from two lines previous in the same current pixel value to obtain a difference value field. These neighboring pixels have the same to be quantized. color subcarrier phasing as the current pixel and will therefore be highly correlated.. The two pixel Figure 1 shows a "nonadaptive predictor" values are averaged to produce the prediction of (NAP) value being subtracted from the current pixel the current pixel value. At this point the algo- value along with the predicted value, PV. The rithm differs from standard DPCM. vlhere the pre- function of the NAP is to further improve the dicted value would simply be subtracted from the prediction of the current pixel. The nonadaptive 2 predictor estimates the difference value obtained The quantizer shown in Fig. 1 has 13 levels. when the DPCM prediction is subtracted from the Each level has a quantization value associated current pixel value (PIX - PV). The subtraction with a range of difference values as indicated in of the NAP value from PIX - PV causes the result- Table I. The quantizer is nonuniform so that more ing difference (DIF) value to be close to zero. levels are provided for small magnitude differ- The smaller the DIF, the more efficiently the quan- ences which would result from subtle changes in tized pixel information can be rransnitted aue to picture content. The human eye is sensitive to the use of Huffman coding prior to rransmission small variations in smooth regions of an image and over the channel. (Huffman coding assigns varia- can tolerate larger variations near transition ble length codewords based upon probability of boundaries where large difference values are occurrence. The application of Huffman coding to more likely to occur. The nonadaptive predictor this algorithm will be discussed later.) discussed previously, acts to reduce the difference values thus improving image quality by reducing the The development of the nonadaptive predictor quantization error. This is because the nonuniform was predicated on the likelihood that the differ- quantizer results in lower quantization error for ence values of adjacent pixels are similar. The small magnitude differences than for large magnitude difference between the current pixel value and its differences. The number of quantization levels, prediction, PV, is estimated and subtracted off by the corresponding difference value ranges, and the way of the NAP prior to quantiztion. The estimate specific quantization values shown in Table I were is simply based on the value of DIF for the previ- experimentally derived through subjective evalua- ous pixel. The NAP is nonadaptive because the tion of sample images processed by computer simula- estimates are prestored and do not change with dif- tion of the encoding algorithm. fering picture content. These prestored values were generated from statistics of numerous televi- The final major aspect of the encoding algo- sion images covering a wide range of picture rithm is the multilevel Huffman coding process. content. The NAP values represent the average dif- Huffman coding of the quantized data allows shorter ference values calculated within the boundaries of codewords to be assigned to quantized pixels having the difference values for each quantization level. the highest probability of occurrence. A separate Table I shows the NAP values corresponding to each set of Huffman codes has been generated for each of quantization level. To give an example using the the 13 quantization levels. The matrix of code values in Table I; if the difference value (DIF) sets is used to reduce the number of data bits for the previous pixel was 40, corresponding to required to transmit a given pixel. The particular quantization level 11, the value of NAP to be sub- Huffman code set used for a given quantized pixel tracted off from the current pixel difference would is determined by the quantization level of the pre- be 38. To reconstruct the pixel, the decoder uses vious pixel (i.e.. if the difference value for the a lookup table to add back in the appropriate NAP previous pixel resulted in quantization level 4 value based upon knowledge of the quantization being selected for that pixel, then the Huffman level from the previously decoded pixel.