Abstract
Gabor wavelets have been successfully applied in various areas from image processing to pattern recognition. This success is mostly due to the fact that Gabor wavelets are equipped with a multi-channel filter capability such that a desired representation of the filter banks can be developed through selecting the parameters, i.e., the center frequency and the orientation of the filter. Although many types of image recognition based on Gabor wavelets have been studied and reported, there are still many valuable topics to be further investigated.The conventional image recognition based on Gabor wavelets is focusing on feature extraction in order to reduce the dimension of the training images. However, this feature extraction is only one of the processes in pattern recognition. For the recognition of the infrared image and optical image, it includes image representation, feature extraction or dimension reduction, classification technique, and the object character. The aim of this dissertation is to construct an image recognition system based on Gabor wavelets for infrared image and optical image. Therefore, An enhanced representation of the Gabor wavelets is proposed, in which the properties of Gaussian mask in Gabor wavelets is developed to enhance the enveloped function, and simultaneously the parameters of the filter based on Gabor wavelets is designed depending on the frequency response of the training images.In addition, the classification technique is important in the pattern recognition. In this dissertation, two classification methods are developed. The modified K-nearest neighbor rule is employed, in which the proposed modified rule combined with variance of the training images is used to find the optimal K value of the nearest distance between training images and testing images. For the neural network classification model, the self-adaptive radial basis function networks is employed, in which the self-adaptive networks has the property that during the training iteration the number of hidden neurons can be either increased or decreased according to the approximation error to prevent over fitting or under fitting.Furthermore, the practical application is also an important issue on the infrared image and optical image. Almost all of the reported Gabor-based image recognitions are applied on the face image. In order to consider the different kind of image, this dissertation proposes an image recognition method based on the Gabor wavelets, which can be applied on both the infrared image and optical image. Some experiments including infrared image and optical image recognitions are given. The good performances are verified through using the proposed scheme in this dissertation.