Pattern recognition of objects contained in images through neural networks

Abstract

This work developed a computational vision system simulator, using artificial neural networks for the geometric shapes classification (square, circle, triangle and rectangle) contained in bidimensional images. The images for the tests were acquired from two capture devices: a scanner and a CCD camera. The images recognition task was divided in two phases: pre-processing and classification through a neural network. In the pre-processing stage, the images were processed by a border detection algorithm, using the Sobel method, that eliminated the image background, just leaving the contour of the object to be recognized. The images acquired by the CCD camera suffered binarization before the processing by the border detection algorithm. The resulting binary images were, then, processed by the Fourier Log-polar transform, to make the system invariant to translation, rotation and scale effects. The results of the log-polar algorithm were the inputs for the neural network. In the classification phase, a probabilistic neural network was adopted, using the Matlab software. Two samples of the processed images were used to train the network, and the remaining ones, for the classification tests. The network classified correctly all the images. In the case of the images acquired by the camera, the number of samples in the training vector had to be increased. The results of this work demonstrated that the neural networks can be used as efficient tools in the geometric shapes recognition task.