Estudio, diseño y optimización de algoritmos para la aplicación de técnicas de aprendizaje estadístico al procesado digital de imágenes

Abstract

This thesis starts on the hypothesis that certain statistical learning tools such as support vector machines (SVM) or RBF neural networks can be applied in low-level image processing tasks. These processing tasks are the rst in a chain basic scheme of image processing. Statistical processing tools have been used in high-level tasks (pattern recognition) with excellent results but its potential has not been fully studied in low-level tasks. Specifically, in this thesis a study of the tasks of edge detection, color images segmentation and impulsive noise removal is made, trying to define alternatives to existing techniques, but based on the use of SVM. Throughout it is shown that alternative proposals can meet or exceed the established techniques with the added advantage of flexibility and adaptation achieved by learning techniques based on synthetic images and easiness of SVM training. While processing tasks discussed in this thesis may seem diverse, they have in common the ability to define classification and regression schemes in which you can apply statistical learning algorithms. Specifically, in the edge detection it is needed to classify the image pixels between those that belong to an edge and those outside, on the color segmentation it is needed to classify the pixels that belong to one color and those that do not and in impulsive noise removal it is necessary to detect the pixels that are noisy and can be reconstructed using the regression. In this thesis have been addressed, therefore, the following tasks for each of the processing techniques studied: -In the edge detection, it was performed an initial state of the art survey to understand the most common detection schemes, their advantages and disadvantages. Subsequently, there was dened an edge detection scheme based on pixel classification by SVM. In this scheme it has been given enough importance to the definition of training, which is done using synthetically generated images. It has made, also, a study of the best training parameters including kernel of SVM. Following the detection scheme, we have defined techniques for obtaining a gradient image from the classification data, and finally, we have defined schemes to mark pixels adapting some existing techniques and proposed some novel solutions. Finally, performance tests have been performed both visually and with objective measures, including some in which the proposed detector is applied to noisy images. -In color segmentation, we have analyzed various existing techniques to understand the points where they can be improved. After this analysis, it is proposed to detect the colors of the image with a simple scheme of detection in the RGB color space. This detection scheme has been compared with existing ones on a particular task, the recognition of traffic signs, demonstrating that the proposed technique can be a feasible alternative and that by its use with lookup tables the performance improvement allows real time processing. -In the section on impulse noise detection, we have studied several existing techniques with different application schemas. Among them was chosen the detection scheme with subsequent reconstruction, due to good results. For detection, we have defined synthetic training images and tested the SVM and RBF neural networks as potential classifiers. Their success in classification task, and the ease of generation of training allowsn to ensure that this is a good alternative to other detection techniques. In the area of reconstruction, median filters have been tested, but modified to improve their performance, and regression SVM schemes, based on synthetic training again. Quality results obtained with dierent measures, allow us to state that the techniques proposed here are an alternative to existing ones and, depending on the images and noise levels, may provide better results.