Object recognition using genetic algorithms with a Hopfield's neural model

Abstract

This paper proposes a hybrid pattern recognition system based on a Genetic Algorithm (GA) with a Hopfield (HP) neural model that can even recognize patterns deformed by the transformation caused by rotation, scaling, or translation, singly or in combination. The proposed method rests on a polygonal approximation technique, which extracts appropriate feature vectors of specified dimensions characterizing a given shape. The features are utilized as inputs in HP neural model type classifier for shape recognition. Object recognition is formulated as matching a global model graph with an input scene graph representing either a single object or several overlapping objects. One can define a matrix for combing the state of each neuron in the HP neural model and consider the matrix as genes of a GA, where the neural nodes constitute the possible matches between the global and scene graphs, and the linkages between the neural nodes comprise the constraints. The present approach combines the characteristics of a GA and a HP neural model, thereby overcoming certain shortcomings, just as initialization and crossover in a GA and setting up the initial value and parameters of the energy function in a HP neural model. It is expected that the qualified convergence to the solution is due to a HP neural model to which information of the gene in the GA is supplied.