Pattern RecognitionLetters 14 (1993) 861-868 November 1993 North-Holland PATREC 1124 A performance analysis of an associative system for image classification Alberto Diaspro Institute of Biophysics, University of Genoa, Genoa, Italy Giancarlo Parodi, Rodolfo Zunino DIBE, Department of Biophysical and Electronic Engineering, University of Genoa, Genoa, Italy Received 27 October 1992 Abstract Diaspro, A., G. Parodi and R. Zunino, A performance analysis of an associative system for image classification, Pattern Recognition Letters 14 ( 1993 ) 861-868. Noise-like coding Associative Memories are applied to image classification. After describing the theoretical frame- work and the imaging architecture, an experimental analysis assesses the system's resistance against increasing noise. Results confirm that the associative 'graceful degradation' provides the classifier with notable noise- insensitivity. Keywords. Associative memory, image classification, robust computer vision, performance analysis I. Introduction In Associative Memory modelling, a datum ('item') is stored by correlating it with other infor- mation ('key') that must be used in further address- ing. This basic principle of data storage and retrieval enables Associative Memories to be content-address- able; distributed data representation makes the memory behaviour resistant when the storing device is damaged. Moreover, a sort of noise-insensitivity can be verified when the information used in ad- dressing is not completely correct. In this paper, the noise-like coding model of asso- ciative memory proposed by Bottini ( 1980, 1988 ) is Correspondence to: Giancarlo Parodi, c/o DIBE, University of Genoa, Via all'Opera Pia 1la, 16145Genova, Italy. used as the basis for the architecture of an image clas- sification system. In comparison with neural models (e.g., Rumelhart and McClelland (1986)), which represent the typical associative approach to pattern classification, Associative Memories exhibit higher structural flexibility and allow easier control. This paper investigates how to exploit the content- addressability and noise-insensitivity of Associative Memories to build up an image-processing system. The resulting image classification does not rely on classic feature analysis or on the maintenance of neural structures. The core of this research is an anal- ysis of the limits of such methodology in terms of classification capabilities. Results confirm that high- level noise can be introduced into the system without loss in classification performance. Section 2 presents a general analysis of the associ- 0167-8655/93/$06.00 © 1993-- Elsevier SciencePublishers B.V. All rights reserved 861