Image recognition by FFT, artificial intelligence and k-nearest neighbors approach

Abstract

This paper presents procedures of processing and recognition of color images by means of artificial intelligence and computer aided training. Two stages of processing have been applied with regard to extraction of features. Processing involves conversion of RGB images in grayscale and binary; and conversion of RGB, grayscale and binary into index formats. Spectral analysis is also applied in here with Fourier’s algorithm for fast conversion concerning indicated graphic formats and segmentation of the real and imaginary complex parts. Activities related with training, assessment and synthesis of feed-forward neural networks and probabilistic neural networks, based on Accuracy and Mean-Squared Error analysis. Special emphasis is made on a method for computer aided training k-nearest neighbors with Euclidean, Euclidean squared, Cityblock and Chebychev metric distances in relation to Cross-validation accuracy value. Some positive indications have been achieved in recognition of assigned graphic images by means of the employed mathematical apparatus.