| Computer vision is a new study widely used in many fields. It is classified into three fields based on Marr theory of computer vision. The extraction and detection of the low level vision feature is important in the system of the computer vision and is especially valuable in practical application. According to the study of the visual physiology,the receptive field of gauglion cell are with outstanding function in the low level vision feature detection such as for the edge ,line , end-stopped ,corner and direction ,etc .Since the lack of the study on the function of the extraction and detection of the low level vision factors, like the line, end-stopped and corner, etc, especially on the precise analysis in scale-space. This research have done deeply study on the mathematical model of the receptive field of gauglion cell——difference of Gaussian( DOG) for its function in the low level vision factors detection. After analyzing the theories significance and application significance of the extraction and detection of the low level vision factors and the latest research of DOG model, the author ascertained the aim and the main content of the research. In the part of the essential characteristic analysis of DOG in this paper, firstly the essential characteristic of Gaussian is analyzed, based on which the essential characteristic of DOG is analyzed both in spatial domain and in frequent domainAfter the analysis of the essential characteristic of Gauss to composing DOG on and the essential characteristic of DOG both in spatial domain and frequency domain, the results show that DOG has good local analysis function both in spatial domain and frequency domain. Theoretically, DOG can be the mother wavelet of the continuous wavelet transform. The realization of the continuous wavelet transform based on DOG is studied and the multiscale analysis function of this transform is proved. In this paper, it is infered that DOG will converge to Gauss and the second order derivative of Gaussian-laplacian respectively on the two extreme conditions of the arbitrarily change of the two scale parameters . After the quantitative analysis and the comparison of the scale detection performance of Gauss function, the second order derivative of Gaussian-laplacian and DOG, it is showed that DOG function has better scale detection performance. The eliminating methods of over-scale edge effection has been studied. The construction of the standard scale function model has beendiscussed.In the part of practical application of the low level vision factors detection of DOG in this paper, the function of scale signal detection in one dimension of DOG is extended, based on which that function in two dimensions is studied. Then the line detection mechanism of Gaussian, Gaussian-laplacian and DOG is analyzed and compared, based on which a practical algorithm of line detection is given. Furthermore an approach to simulating the location, orientation and shape of corner is proposed by detecting two maxima in different scale. And the result of experiment with signal in strong noise is given by the detection of DOG. Based on the mechanism of line detection and corner detection of DOG, a new characteristic scale and multiscale analysis-based texture segmentation algorithm is proposed and some experiment result are given.In the part of the function of DOG in analysis of image and visual phenomena in this paper, a "point model" method is proposed for the complex signal detection in two dimensions such as endstopped detection and scale point detection. Furthermore the responses of all kinds of scale signal to DOG is compared and summarized. Then it is proposed that the receptive field of gauglion cell has "attention" function in visual system. Based on the "attention" function of the receptive field of gauglion cell and combined with the practical image processing of DOG, some visual phenomena and illusions are analyzed and explained such as excalmatory mark, the function of texture in visual information proce...
|
PDF Full Text Request