Visual Recognition of Simple Objects by a Connection Network

A difficult problem in vision research is specifying how meaningful objects are recognized using the visual feature information extracted from an image. The fundamental issue involves the interaction of different levels of representation of visual information. The technical and theoretical problems that must be addressed in specifying this interaction arise in any attempt to model visual object perception. We attempt to deal with some difficult aspects of this process within the context of Feldman's Four Frames model of visual perception [Feldman, 1984]. The model consists of four continually interacting representational frames, expressed in terms of a massively parallel, connectionist formalism. Within the Four Frames model, the problem of accessing object representations using visual feature information can be defined in specific computational terms. We present in this paper the detailed design of a connectionist model as a possible solution to some of the major problems in the visual recognition of objects. The model proposes that an object is represented as a hierarchical structure of geometric subparts. Recognition proceeds by determining in parallel that all subparts of an object are present in the image, and then sequentially verifying that each subpart is in the proper spatial relation to the others. Implementation results demonstrate that the model can recognize any of a set of simple objects given fairly general feature input Although the model is developed in the context of a drastically simplified visual domain, the principles it embodies are argued to adhere to many of the behavioral and biological constraints of real-world vision