Feature Point Extraction In Gabor Based Energy Space

Image feature extraction is a fundamental problem in computer vision, which is also the foundation and main component of image matching, object recognition, video analysis, image retrieval and so on. Since the concept of image feature was proposed, many remarkable achievements have been made in this field. Moreover, many research findings in physiology, neurology and many other subjects inspire the study in this field, especially many methods, which imitate biological visual processes, achieve great success in settle real problems. However, the image feature detection algorithms has many defects and limitations in many areas, such as the reference models limit the types of features could be detected, the complicated process of computing go against the real-time of the vision systems. The more complicated tasks of vision systems, the bigger changes of images caused by environment and the more visual information to be extracted demand the feature extraction to be effective for more image structures. The most previous feature detectors are only used to extract some special features, such as the Robert, Sobel, Canny operator in edge detection, Harris and improved methods in corner detection and the SIFT and SURF methods widely used in recent years. These methods, which are designed for special structures, have limits in the other structures and are difficult to achieve convergence between features. Therefore, developing a detection method from a new perspective, which are applicable to most structures, is the focus of the research with difficulty.In this paper an energy based feature extraction is proposed for a variety of image structures with stability for illumination change. Meanwhile, it can achieve a large-scale range of feature detection with a relatively simple implementation. The main contents are as follows:1. The Gabor wavelet is treated as the tool for the image multi-scale analysis after a review and comparison between the Gabor and Log-Gabor function. Then a detailed analysis of the types of image features is made from the viewpoint of phase. After an effective classification of image features, we research the energy based feature extraction.2. The energy space of image is constructed by the Gabor wavelets, which makes it possible to extract features.from a larger range of scales. And the change of features along the scale is studied based on the new definition of characteristics scale related to energy. Finally, a recursive form of convolution formula is derived to achieve a rapid construction of energy space.3. After analyzing the relation between image features and image energy function, an energy-based feature detection is proposed. An iterative search method is adopted to find the extreme points, which represent the spatial position and characteristic scale in Gabor based energy space. This method works in spatial-frequency domain, which is not restricted by the reference model, which makes this detector be more appropriate to real image and increase the capacity of resisting disturbance of vision systems.4. In view of the poor real-time performance of feature detection, this paper discusses the Gabor-based characteristic scale selection, and proposes a fast feature detection method combined with improved Fast-Hessian method, which could detect feature points with a wide range of scales. The experiments results show the feasibility and adaptability of the method.