A MULTI-LAYER ANALOG VLSI ARCHITECTURE FOR TEXTURE ANALYSIS ISOMORPHIC TO CORTICAL CELLS IN MAMMALIAN VISUAL SYSTEM
In Digital as well as Analog technology VLSI Artificial Neural Networks (ANNs) play very important role in recent days. In this study researchers present an analog VLSI architecture for texture discrimination. The outline of this architecture is derived on the basis of a formal neuron model of cortical cells which are mapped directly into silicon with high efficiency.
Cortical cells acts as texton detectors as they respond optimally to oriented elongated blobs and their terminators to oriented gratings of appropriate spatial frequency [1].The properties of a cell can be characterized by its response to specific visual stimuli. Receptive field of the neuron is a area of the visual field where stimuli bring out the neural response. 2D spatial response is one of the receptive field which is known by receptive field profile. 2D spatial response profile is a distribution of neuron sensitivity to visual stimulation. More specifically, the linear 2D response profile refers to a spatial weighting function by which a local region of the image is multiplied and integrated to generate that neuron’s response.
Researcher’s main aim is to visualize and specify an analog VLSI neural system inspired by the structure of the visual system The basic outline of the analog architecture is composed of two layers of computational units: the first one is the input-and-preprocessing layer, the second one is the processing layer. The first layer i.e. input layer is the first layer in chain of visual processing; the input image is pre-processed with operations similar to those accomplished by the retina [2]. Generally input layer is organized with a layer of mean extractors. The pre-processing layer is used for compensation for different illumination conditions. This operation can be realized with elements similar to those used in the subsequent process, because also this operation can be realized with a convolution operation. The processing layer is shown in fig 1 (a). Neighboring units are connected with different areas of the input layer by the spatial organization of their receptive fields and this is analyzed by a complete set of orientations i.e. hypercolumn which is shown fig 1(b).
References:
[1.] Albus, K., “A Quantitative Study of the Projection Area of the Central and Paracentral Visual Field in Area 17 of the Cat”, Exp . Brain Res., vol. 24, pp. 159–202, 1975.
[2.] Bertero, M., Poggio, T. and Torre, V., “Ill-Posed Problems in Early Vision”, Proceeding of the IEEE, vol. 76, pp. 869–889, 1988.