Preattentive computer vision towards a two-stage computer vision system for the extraction of qualitative descriptors and the cues for focus of attention

摘要

A consensus is emerging from current trends in machine vision (MV) research indicating that a two-stage processing system which imitates natural vision may be the way forward in developing efficient and generic MV systems. This paper reviews previous trends in modelling and recent implementations of the components of two-stage vision systems. In addition a new model for the preattentive stage of visual processing is developed. The modelling process for this stage is novel in that it is guided, not by the neuroantomy of natural vision systems, but by their principles of operation. Such preattentive mechanisms cannot be accessed directly by physical examination, they can only be infered by psychophysical experiment. These indicate that true preattentive processing exhibits execution times which depend only upon the size of the input image, and are independent of the quality or type of image data. Application of this constraint to the design of a preattentive MV system has significant implications and these are discussed. The proposed model is significantly different from previous suggested models in that it is capable of handling images of varying complexity containing multiple objects of varying sizes whilst conforming to the principle of data independent processing times. Also reported is a case study which uses the model to replicate an industrial inspection task currently undertaken by skilled operators, namely the grading of fine powders with respect to the size and abrasive properties of the constituent particles. This particular application is well suited to the study of preattentive visual processes as it is a good example of the innate abilities possessed by human beings to perform complex visual categorization tasks for which no unique analytical representation exists, and for which reaction times do not depend upon the complexity of the image. The case study highlights the considerable potential and relative simplicity of the proposed methodology. Real image data containing multiple objects at various scales are used. It is demonstrated that various qualitative descriptors, perceptual groupings and conspicuity of locations can all be made simultaneously available at the preattentive stage. However, application of the model to a single visual inspection task leaves many issues still to be addressed. The work includes a discussion of the various remaining open questions with the aim of motivating further research into generic two stage MV systems which have broad ranges of applicability.