IRWIN AND JOAN JACOBS CENTER FOR COMMUNICATION AND INFORMATION TECHNOLOGIES Automatic Gain Control of Images Motivated by Human Vision

Biologically-motivated Automatic Gain Control (AGC) scheme of processing along various image dimensions is presented. Images are represented in a multidimensional space that incorporates less investigated dimensions such as curvature, size, depth, convexity/concavity and more. Similarly to the effect of AGC on processing of intensity in vision, the proposed scheme for multidimensional image processing and computer vision enhances and emphasizes important image attributes along each dimension of the representation by adaptive nonlinear filtering. This achieves vision-like intelligent image processing and new powerful means for computer vision. The proposed AGC scheme is analyzed for its SNR characteristics, and for its action on several specific inputs. The results are compared with effects known from visual psychophysics, exhibiting reproduction of visual illusions. Finally, examples of applications in computer vision are presented. These include processing of HDR (High Dynamic Range) images, enhanced edge detection and interpolation of curves partly occluded during the acquisition process. Incorporating a generic neural network AGC scheme along all the visual dimensions, constitutes a universal, parsimonious and unified model that can span a multidimensional HDR, and yet be sensitive to fine details along all the dimensions. Implementation of the multidimensional AGC in image processing and computer vision may also contribute to the development of a metric for image space, and facilitate further development of new means for recognition and classification.