Acknowledgments: I would like to thank D. Marcuse for several calculations and numerous helpful discussions relating to this work. Conversations with P. Henry and G. Williams were very helpful. A. Gnauck provided the fast rise time drive signal used in these measurements. R. A. LINKE 25th April 1984 AT&T Bell Laboratories Crawford Hill Laboratory Holmdel, NJ 07733, USA References 1 MALYON, D. j., and MCDONNA, A. P.: '102 km unrepeatered mono- mode fibre system experiment at 140 Mbit/s with an injection- locked 1-52 /an laser transmitter', Electron. Lett., 1982, 18, pp. 445-447 2 ICHIHASHI, Y., NAGAI, H., MIYA, T., and MiYAJiMN, Y.: 'Transmission experiment over 134 km of single-mode fiber at 445-8 Mb/s'. Post- deadline paper, IOOC83, Tokyo, June 1983 3 KASPER, B. L., LINKE, R. A., CAMPBELL, J. C, DENTA1, A. G., VODHANEL, R. s., HENRY, P. s., KAMINOW, i. p., and KO, J.-S.: 'A 161-5 km transmission experiment at 420 Mb/s'. Postdeadline paper, ECOC 83, Geneva, Oct. 1983 4 LINKE, R. A., KASPER, B. L., CAMPBELL, J. C , DENTAI, A. G., a n d KAMINOW, i. P.: 'A 1 Gb/s lightwave transmission experiment over 120 km using a heterojunction APD receiver'. Postdeadline paper WJ7, Conference on optical fiber communication, New Orleans, Jan. 1984 5 IWASHITA, K., NAKAGAWA, K., NAKANO, Y., a n d SUZUKI, Y.: 'Chirp pulse transmission through a single-mode fibre', Electron. Lett., 1982,18, p. 873 6 LiN, c, LEE, T. p., and BURRUS, c. A.: 'Picosecond frequency chirp- ing and dynamic line broadening in InGaAsP injection lasers under fast excitation', Appl. Phys. Lett., 1983, 42, p. 141 7 OLSSON, N. A., DUTTA, N. K., and LIOU, K.-Y. : 'Dynamic linewidth of amplitude-modulated single-longitudinal-mode semiconductor lasers operating at 1-5 /<m wavelength', Electron. Lett., 1984, 20, p. 121 8 KRESSEL, H., and BUTLER, j. K.: 'Semiconductor lasers and hetero- junction LEDs' (Academic Press, 1977), Chap. 17 9 TSANG, w. T., OLSSON, N. A., and LOGAN, R. A.: 'High-speed direct single-frequency modulation with large tuning rate and frequency excursion in cleaved-coupled-cavity semiconductor lasers', Appl. Phys. Lett., 1983, 42, p. 650 10 KAMINOW, i. p., KO, J.-s., LINKE, R. A., and STULTZ, L. w.: 'High- speed 1-55 /im single-longitudinal mode ridge waveguide laser', Electron. Lett., 1983, 19, pp. 784-785 11 THOMPSON, G. H. B.: 'A theory for filamentation in semiconductor lasers including the dependence of dielectric constant on injected carrier density', Opto-Electron., 1972, 4, pp. 257-310 12 FURUYA, K., SUEMATSU, Y., and HONG, T.: 'Reduction of resonance like peak in direct modulation due to carrier diffusion in injection laser', Appl. Opt., 1978, 17, pp. 1949-1952 13 MIDWINTER, J. E.: 'A study of intersymbol interference and trans- mission medium instability for an optical fiber system', Opt. & Quantum Electron., 1977, 9, pp. 299-304 NEW COMPLETE PROCESSING SYSTEM FOR RECOGNITION AND CLASSIFICATION OF OBJECTS IN MULTIPLE IMAGE FRAMES Indexing term: Image processing A complete and efficient set of procedures is proposed to solve problems related to object detection and classification in the processing of a sequence of image frames. Nonlinear smoothing, adaptive thresholding and classification by iner- tial invariants are proposed. Introduction: Image processing has been of growing interest in recent years in many applications. Two main steps can be identified in any image processing system: 'preprocessing' and 'processing'. Preprocessing deals with eliminating possible 474 error causes in the subsequent processing step. Processing depends by what we aim at, and can deal with feature extrac- tion, image transformations, image segmentation, object iden- tification etc. In any case it can be a logical sequence of processing substeps. If we are dealing with sequences of frames (e.g. in moving objects analysis), a third processing step, 'tracking', is needed to make predictions and track the objects in the frames. Qualification of the previous processing steps for consecu- tive frames analysis is Preprocessing: noise reduction Processing: frame segmentation, object classification, object identification Tracking: object tracking. In the following we propose a new and very efficient set of algorithms to solve problems related to the first two steps as above. Noise-reduction algorithm: Using spatial masks to consider spatial correlation of each pixel is widely covered in the liter- ature. 1 " 4 Anyway its processing requests often pose objective limits to its implementation on on-line systems. A five-order moving average, moving along a preferential direction, line by line, according to the following recursive algorithm, is pro- posed where x, is the original grey level in the fth position and x, the new one: / 6 [3, 253] x, =x, x 2 =(x, +x 2 )/2 X 3 X 5 )/5 = ( X 2 (1) Five-order has been proved to give nonrelevant nonsym- metrical deformations with an interesting computational saving with respect to the other methods previously indicated. Segmentation-classijication-recognition: Segmentation involves object-background separation and single-object detection. The mean value for a fixed neighbourhood of the absolute minimum (to eliminate residual noise spikes) is evaluated on the pixels for each line of the frame. A symmetrical band is established over the mean value of the line (its width is chosen according to a look-up table experimentally defined), to separate grey levels belonging to the background from the object's ones. Single-object identification is hence obtained: in the first line the object's pixels are labelled with different iden- tifiers if nonadjacent; in subsequent lines adjacent pixels are labelled by the maximum identifier of the adjacent pixels in the previous line; a final processing is hence required to solve possible ambiguities and to connect with the same label pixels belonging to the same object. Classification and recognition are proposed, based on iner- tial invariants and a matching procedure with an appropriate 'object models space'. This, differently from previously pro- posed algorithms, 1 ' 5 allows some variations in the actual objects against the models, like scaling, rotation etc., without introducing errors in the recognition phase. In particular the area, the centroid position, the maximum elongation axis and the 'elongatedness' defined according to expr. 2, are used to build a suitable set of state variables for each object under analysis: Elongatedness ^ (A 2 — A,)/(A, + A 2 ) (2) where A 2 > A t are eigenvalues of and n pq is the central moment of order pq. Sensible reduction in the cardinality of the set of the object models is obtained, against other identification methods pro- posed in the current literature. ELECTRONICS LETTERS 24th May 1984 Vol. 20 No. 11