Architecture of the parallel-hierarchical network for fast image recognition

The paper presents a forecasting method based on the parallel-hierarchical (PH) network and hyperbolic smoothing of empirical data. The multistage approach to image processing includes main types of cortical multistage convergence. One of those types occurs within each visual pathway and another one between the pathways. This approach maps input images into a flexible hierarchy, which reflects the complexity of image data. Procedures of the temporal image decomposition and hierarchy formation are presented as mathematical terms. The multistage system highlights spatial regularities, which are passed through a number of transformational levels to generate a coded representation of the image which encapsulates, in a computer manner, a structure on different hierarchical levels in the image. At each processing stage, a single output result is computed allowing rapid response from the system. The result was presented as an activity pattern, which can be compared with previously computed patterns on the basis of the closest match. An average prediction error is 0.55% for the developed method, and 1.62% for neural networks. It was pointed out the developed method is more efficient for the real-time systems comparing to traditional neural networks in forecasting energy center positions of laser beam spot images.

Author: Andrzej Kotyra
Conference: Title