Object recognition by saccadic parts verification

A model for the sequential processing of image 'parts' alias subobjects for the purpose of identifying visual objects is presented. The implementation uses three different coherently working neural networks to accomplish the task. One is for coarse resolution hypothesizing, one for verification via fine resolution subobject identification, and the third is a sequentially processing saccade generation net. The mathematical/architectural model is implemented on a UNIX computer with MOTIF interface. In addition a comprehensive set of positional, scale and rotational invariance (PSRI) conditions are tested using CCD camera inputs of real world objects. The data collected from experiments describe the system to have 100% PSRI for real photographs. Samples over a wide range of other real world conditions such as varied illumination and cluttered scenes are also correctly recognized.<
>