TURBO CODE APPLICATION TO IMAGE TRANSMISSION

C. Tanriover*, P. Chippendale*, B. Honary*

Abstract This paper explores the application of the error resilient coding scheme, Absolutely addressed Picture ELement coding (APEL), to image transmission over noisy radio channels. To improve the performance of APEL at low signal-to-noise ratios, turbo coding is introduced into the system. Demonstrated through Gaussian and Rayleigh fading channel simulations, this novel combination will be shown to correct and restrict the propagation of the majority of errors incurred during transmission. 1. APEL Image Coding APEL [1,2] is a loss-less, robust image coding system which translates variable sized pixel areas of pre-defined dimensions into independent picture blocks (pels). Each pel is issued with two co-ordinates, x and y, establishing an absolute location with respect to an origin. As the underlying APEL coding technique operates on a binary level, the encoding of grey-scale or colour images employs a Bit Plane Coding (BPC) [3] stage. The BPC stage furnishes the APEL encoder with a colour coding sequence to represent a given source image in binary planes. Taking each extrapolated binary plane in turn, a recognition algorithm searches through each image looking for square areas of black pixels; starting with large square pels during the first scan, then repeating this process in multiple passes selecting pels of decreasing magnitude. The maximum size of the initial pel is limited according to the anticipated nature of the channel, consequently less information is lost should corruption occur. Once all of the square pels of an efficient size have been removed from the plane, run-lengths of various geometries are used to encode the residue. Fig. 1 illustrates an APEL encoded section of a grey-scale image. Here, it can be seen how (x,y) co-ordinates are assigned to pels of various geometries.

X
Horiz Run = 2, (6,9) Square = 4, (8,6) Vert Run = 4, (2,4) Square = 5, (3,4) Horiz Run = 7, (3,3) Horiz Run = 4, (2,2) Origin

Fig. 1 APEL coded section of an image The data-stream created from this process can be pictured as a succession of (x,y) addresses, grouped according to the same size they represent and interspersed with control symbols (Fig. 2). These symbols not only serve to provide synchronisation markers, but in addition convey pel geometry metrics to the decoder.
Control Symbol (x1,y1) (x2,y2) (x3,y3) (x4,y4) ........ (xn,yn) Control Symbol

Absolute addresses

Pel parameter and synchronisation information

Fig. 2 Breakdown of APEL data stream

*

Department of Communication Systems, Lancaster University

The APEL absolute addressing scheme alleviates the need for End Of Line (EOL) symbols and, as each codeword is independent, offers a solution to the problems of horizontal and vertical error propagation. Additionally, as each pel has its own address, it is possible to interleave them within the transmitted data-stream. This versatility can be utilised in many ways, for example: pels pertaining to important image detail can either be dispersed throughout the data-stream or transmitted at the start depending on channel conditions or operator preference. 2. Application of Turbo Coding By employing iterative soft decoding principles, turbo codes [4,5] achieve bit error rates close to the Shannon limit. This powerful forward error correction technique has been applied to APEL coding to provide a robust image communication means (Fig. 3).
Source Image Decoded Image

COMMUNICATION CHANNEL
APEL Encoder Turbo Encoder Transmitter Receiver Turbo Decoder APEL Decoder

Fig. 3 Turbo-APEL serial concatenation In this study, a rate recursive systematic convolutional code with constraint length 3 is used as the component of a turbo coding scheme. The encoder employs parallel concatenation as illustrated in Fig. 4. The binary information to be encoded is represented as uk, with ck,1 and ck,2 signifying the parity bits of the first and second component encoders, and where D stands for a bit delay. Symbol represents pseudo-random interleaver, which accepts blocks of 8000 information bits.

uk

uk ck,1
D D puncturing

ck,j

ck,2

Fig. 4 Turbo encoder block diagram Binary information, uk, is fed to the component encoders in order to generate the associated parity bits, ck,1 and ck,2, which are selected alternately by the puncturing block. In other words, for a binary input sequence of u1, u2, u3, u4, the encoded sequence would be u1, c1,1, u2, c2,2, u3, c3,1, u4, c4,2. The fact that the input sequence uk is retained at the output, is due to the systematic nature of the encoder. Fig. 5 shows the turbo decoder which contains two component decoders concatenated in parallel, performing the suboptimal log-MAP algorithm. Lek,i is the extrinsic information related to the kth information symbol, rk, provided by the ith decoder, and yk,i is the parity symbol associated with rk, and fed into the ith component decoder. rk yk,1 Lc component decoder 1

Lek,2
hard decision

Lek,1
component decoder 2

-1
r`k

yk,2

Fig. 5 Turbo decoder block diagram

In order to perform the MAP algorithm, an accurate estimation of the channel conditions, i.e. the noise variance, is required. Hence, a channel reliability factor, Lc, is fed into the decoder, which equates to 4.a.Eb/No [4]. In this expression, a denotes the fading amplitude, which is 1 for a Gaussian channel and Eb/No is the bit energy to single sided noise spectral density ratio. De-interleaving, -1, is the inverse operation of interleaving and is performed at the output of the second component decoder. When the required number of iterations (fixed at 16 in our case), have been completed, the soft output of the second decoder is de-interleaved and a hard decision is made to obtain rk. 3. Results Simulations over AWGN and Rayleigh channels, at various signal-to-noise ratios, verify the excellent performance of this novel APEL-turbo scheme. Performance comparisons were carried out between turbo coded APEL, JPEG, and BMP image files. To provide a fair comparison between the APEL and JPEG images, compression level had to be the same. This resulted in the data reduction for both APEL and JPEG to be around 5 to 1. In the assessment of bit errors within a received image file, a quantitative pixel error rate offers a fairer comparison. Therefore, it is appropriate to represent the error performance in a ratio called Pixel Error Rate (PER), which is a measure of the degree of image degradation. Through the analysis of the ith received pixels variance from its transmitted value, a measure of visual disturbance, i, can be quantified as in (1), where ti and ri represent the transmitted and received pixel colours respectively, for an n colour image.

i =
From (1) it follows that the PER is calculated as in (2),

ri ti n

(1)

PER =

1 XY

i =0

XY

(2)

where X and Y are the horizontal and the vertical resolution of the image, respectively. 3.1. Gaussian Channel results The 3 plots shown in Fig. 6 illustrate the average results attained from 50 tests performed over a simulated AWGN channel. The number of pixel errors encountered are highlighted from 10 million pixels. As the figure indicates, the performance of the JPEG-turbo scheme is very poor in the 1.0 1.4 dB range. This is due to the inherent fragility of the JPEG structure and its inability to correct, or restrict the propagation of, any errors. This result was not unexpected. The variable length Huffman codewords employed by JPEG do not provide any protection against errors and a single error can lead to catastrophic collapse. The complete opposite of this interdependence can be seen in the BMP file format, and in accordance the performance of the BMP-turbo scheme is good throughout the range. Here, when errors cannot be repaired by the turbo decoder, only pixels with corrupted bits are affected. Finally, performance close to, and at times surpassing, that of the BMP-turbo model is achieved by the turbo coded APEL. In the region after 1.175 dB, the post-processing techniques employed by APEL recover many of the damaged pixels which had simply been inserted in the case of BMP. To observe the visual impact of the pixel errors, samples of the various file formats have been decoded at a signal-tonoise of 1.175 dB (Fig. 7). In this example, the incongruous and clustered nature of the pixel errors in the APEL file, makes it appear to be of similar quality as the BMP one. However, it has to be stressed that the APEL file is a fifth the size of the BMP one. 3.2. Rayleigh Channel results Turbo coded JPEG, bitmap and APEL images were also transmitted over a Rayleigh channel with a maximum of 200 burst errors introduced randomly in each transmitted block. Fig. 8 illustrates the system performance in the presence of burst errors. Due to the severe effects of burst errors, Turbo-JPEG fails to maintain data integrity and synchronisation after decoding (Fig. 8A). Again, the fragile structure of the Huffman codewords makes it almost impossible to withstand such channel conditions. In addition, given that the turbo decoder used in this application is unable to correct burst errors, image transmission with Turbo-JPEG becomes very unreliable in such conditions.

100000

number of pixel errors

75000

EL AP boTur

oTurb G JPE

50000

P BM orb Tu

25000

0 1 1.1 1.2 1.3 1.4

Eb/No (dB )

Fig. 6 Image transmission performance over Gaussian channel

C
Fig. 7 Gaussian channel Results for Eb/No = 1.175 dB A- Turbo coded JPEG, B- Turbo coded APEL, C- Turbo coded BMP

Turbo-APEL (Fig. 8B) performance in the presence of burst errors, is again visually comparable to that of Turbo-BMP. Channel errors which affect pels from various bit planes can be corrected through an analysis of the other planes. In other words, the post-processing techniques introduced by APEL coding provide a powerful means of interpolating pixels using valid image information. However, as the number of burst errors per information block is increased, distortion in APEL images remains noticeable despite the post-processing.

C
Fig. 8 Rayleigh fading channel results for Eb/No = 4.0 dB A- Turbo coded JPEG, B- Turbo coded APEL, C- Turbo coded BMP Unlike the Gaussian channel, the PER plots produced by the more severe Rayleigh case were found to be inconsistent and vary from one transmission to the next. It was also observed that the dynamic range of the decoded image quality was noticable. Hence the use of a different measure for evaluating the quality of the decoded images has been investigated. When evaluating image performance, the most appropriate tool to employ is the human eye, as afterall the recipient of the transmission will be human. Keeping this point in mind, a psychometric study based on the subjective rating of a group of 20 individuals from various backgrounds, was conducted to obtain mathematical data for quantifying pixel disturbance levels. Each individual was shown typical turbo coded APEL and a BMP image files transmitted over the Rayleigh fading channel, and then asked to rate each of them in a scale from 1 to 10 (a higher rating indicates better quality). This procedure was repeated for various channel conditions to generate a range of sample points (see Fig. 9). In the opinion of the test subjects, the BMP images that appear in Fig. 9 are clearer than the APEL ones below 3.7 dB. After this point, APEL image quality is perceivably better than the BMP one by a significant amount. The small perturbations in the curves correlate to specific images where sensitive areas of a particular image are damaged. When

burst errors are concentrated on image detail which is deemed more important, e.g. the cats face, they tend to disturb the human observer more [6]. Consequently, in some cases the perceptual quality of an image can locally decrease despite improved channel conditions. This is exemplified at 5 dB where several APEL errors obscured an area of the cats face. In the case of BMP images, since the pixel errors appear as trails of corruption, the perceptual quality increases more gradually as the channel improves.

10

Perceptual Quality (out of 10)

Turbo coded BMP Turbo coded APEL

0 3 3.5 4 4.5 5 5.5 6 6.5 7

Eb/No (dB)

Fig. 9- Psychometric evaluation of turbo coded BMP and APEL images simulated over Rayleigh Fading Channel 4. Conclusion We have proposed the combination of APEL and turbo coding to provide a reliable, compressed image transmission system at low signal-to-noise ratios. Even though APEL coding is used with a turbo decoder that is not powerful enough to correct heavy burst errors, the interleaving stage within the APEL structure is observed to minimise the visual impact of errors. This visual improvement is achieved through the dissemination of burst errors both the constituent bit-planes and the entire image, thus providing an interleaver gain at the decoder. Whilst the majority of bit errors within the APEL image are corrected via iterative decoding, any which fail to be detected (and thus perhaps be falsely inserted as erroneous pixels) are restricted as a result of the robust data structure. Since the interleaving stage in APEL coding distributes pixel errors across the entire image (Fig. 8B), the resulting small clustered errors that occur in this case can often be less disturbing to the eye than erroneous pixel trails (Fig. 8C), as Fig. 9 suggests. The novel combination of the APEL source/channel image coding technique, and additional channel protection provides not only a resilience to Gaussian type errors, but also offers a powerful tool for the restriction of burst errors. Acknowledgements The authors wish to thank DERA Malvern and NDS Ltd. for their support and use of their equipment and also Psychologist Paolo Deluca, (Lancaster University) for his constructive advice and suggestions. References [1] Chippendale, P., Honary, B., Arthur, P. and Maundrell, M.: International Patent Ref.: PCT GB 98/01877, Data Encoding System [2] Chippendale, P.: Transmission of images over time-varying channels, PhD Thesis, August 1998 [3] McConnell, Bodson and Schaphorst, 1992, FAX : Facsimile Technology and Applications Handbook, ISBN 0 89006 495 5 [4] Berrou, C., Glavieux, A., Thitimajshima, P.: Near Shannon Limit Error Correcting Coding and Decoding: TurboCodes, IEEE Proc. ICC 93 Geneva, Switzerland, May 1993, pp. 1064-1070 [5] Hagenauer, J., Offer, E., Papke, L.: Iterative Decoding of Binary Block and Convolutional Codes, IEEE Transactions on Information Theory, Vol. 42, No. 2, March 1996 [6] Kosslyn M.S., Osherson D.N., Visual Object Recognition, Irving Biederman, Visual Cognition, 1995 Massachusetts Institute of Technology, volume 2, pp. 121-165