Modeling And Image Representation Of Primary Visual Cortex Facing Functional Column

Recognizing an object from its background is always a very challenging task for pattern recognition, especially when the size, pose, and illumination of the object and the background are changing. The most essential method for handling this classical problem is to learn and define the structure of an object using its topological or geometrical features and components.Once we have several positive samples of a type of object, we always draw a prototype for the object. This is an inner representation which needs to be memorized. In a computer program, this requires some kind of formal data structure which can describe that prototype. A direct means is by choosing some kind of knowledge representation methods, such as a semantic network, production rule, frame, or ontology. But from the point of view of knowledge engineering, a good knowledge representation scheme should be able to smoothly join two levels:the level of knowledge being acquired from the learning phase, and the level of knowledge being used in the reasoning phase. The knowledge we are concerned with here pertains to the structural characteristics of an object.However, most of the researchers focus on theories and algorithms proposed from the standpoint of engineering. The underlying physiological principles are ignored and the theories lack a complete computational system and uniform algorithms, which makes it impossible to solve problems of complex scene recognition. By contrast, vision perception system of higher mammalian animals is capable of perception tasks in various complex environments. With local visual information well combined in the brain, mammalian animals produce aggregated and harmonized sensation when observing scenes. They are able to distinguish objects of different sizes and features from highly complex background. The research of visual perception of higher mammals hence makes a possible break through.Here we want to simulate the primary visual cortex for its hub role in visual in formation processing:all information from the lower retina ends here and all information up to the higher cortex starts here. So the information stored in the primary cortex is sufficient enough for upcoming advanced processing. In the V1area of striate cortex many cells respond strongly to special orientations. This kind of directional selectivity was studied and the concept of an orientation column in V1is proposed. The orientation column model inspired us to design a similar array to record the orientation distribution of an image. The outputs of the orientation columns are active chips and we believe that contour information must be included in this distribution. We developed line-context descriptors from the set of active chips and regard them as the representation unit on high-order statistical information of the target. To classify the targets we reorganize the descriptors obtained from several prototypes into a database sheet and using the sheet as the feature representation of object.Here the array of simulating orientation columns in the primary visual cortex is sparser than raw data, and denser than formal concepts. The semantic responsibilities assigned to this array are heavier than raw data and lighter than formal concepts. So it is a very good place, with the right granularity, to let data and concept converge.The experimental results show that our algorithm can learn from a small training set and can recognize the same types of objects under a natural background without any preliminary information. This bio-inspired representation platform is promising for handling semantic-concerned problems that require an intensive definition of semantics.