Computational vision enhancement through genetic selection of biologically inspired receptive field geometry

The paper presents a genetic selection of biologically inspired receptive field geometry to improve pattern recognition in neural network classifiers. A genetic algorithm is employed to select the x, y dimensions and orientation of the receptive fields in a four hidden layer neural network with two planes per layer. Networks were ranked based on the fitness criterion: best generalization performance on a handwritten digit database. Results show strong correlation between the neural network performance and the receptive field x and y dimensions and orientation. The genetic algorithm improves classification performance by selecting appropriate receptive field size and orientation. The best receptive field configuration results outperformed those of perception based models. The proposed method allows a comparison among different architectures of receptive fields to find general patterns of improved performance.