A new art-based neural architecture for pattern classification and image enhancement without prior knowledge

摘要

The visual field is not perceived as an array of independent picture points. Instead, it is usually seen as consisting of a relatively small number of patterns. Whenever a picture is converted from one form to another, e.g. imaged, copied, scanned, transmitted, or displayed, the “quality” of the output picture may be lower than that of the input. In the absence of knowledge about how the given picture was actually degraded, it is difficult to predict in advance how effective a particular enhancement method will be. In this paper, the formulation of a new neural architecture is presented based on adaptive resonance theory (ART), for the pattern classification and image enhancement in the presence of noise without prior knowledge. The underlying theory and the improvement of the ART model are first investigated in classifying optical character patterns. Based upon the result, the two-layer ART model is incorporated into a four-layer neural network which is proposed whereby pre-established generalized enhancement templates are used as region or contour detection exemplars in order to fill in the gaps and eliminate the noise in a pattern without any prior knowledge of the image itself.