Abstract
Image processing and computer vision encompass a broad domain of applications, such as image segmentation for medical image analysis, aerial image segmentation for autonomous navigation, object recognition, biometric recognition, security surveillance,etc. The extraction of features as helpful information for the interpretation of digital image data is an essential task and interesting research subject in image processing and computer vision domain. In this report, image filtering for image low-level feature detection and extraction is investigated, which mainly involves: an objective evaluation of ridge/valley detection techniques, proposing two multiscale line feature detection tech-niques, and proposing a new anisotropic corner detection approach. First, to have an intuitive context and introduction, image filtering basics, basic contour detection techniques using first order and second order operators, and evaluation metrics and tools are overviewed. Meanwhile, the scale space theory and axioms were also investigated in order to find the basic criteria and constraints in the development of multiscale feature detection methods. Secondly, we objectively evaluated the state-of-the-art ridge/valley detection and extraction techniques. The objective analysis of a ridge characterized as a thin and complex image structure is essentially important, for choosing, which parameter values correspond to the suitable configuration to obtain accurate results and optimal performance. The optimal parameter configuration of each filtering technique aimed for the image salient feature analysis tool has been objectively investigated, where each chosen filter’s parameters correspond to the width of the desired ridge or valley. The comparative evaluations and analysis results are reported on both synthetic images, distorted with various types of noises, and thereafter real images. Thirdly, to deal with the multiscale structure of line features, we proposed a new line feature detection and extraction technique, which is based on the second order semi-Gaussian filter. The experiments were ledon both synthetic and real images, and the obtained result demonstrated more optimality to the state-of-the-art line feature detection based on filtering approaches. Next, we proposed another multiscale line feature detection and extraction technique composed ofbi-Gaussian and Semi-Gaussian Derivative Kernel. The proposed filter is able to precisely extract the complex, narrow, and adjoin linear structure and is adapted for multiscale ca-pability. The proposed filter is validated with experiments on different images containing complex adjoin linear structures with different scales. The extracted linear structure on both synthetic and real images has shown to be more efficient than classic linear structure extraction techniques. Regarding the corner detection techniques, we performed first an objective repeatability evaluation of 12 state-of-the-art corner detection based on a filtering approach. There exist different techniques for keypoint detection; as filtering is concerned, we have focussed on direct computation on the gray-level analysis of interest point detection because of its straightforward implementation. Our evaluation as an application to feature matching has been executed in the context of underwater videosequences. Finally, in this work, a new anisotropic corner detection method based on the formulation of causal filtering is proposed. The proposed corner detector arguably performs better in the case of localization precision. The experiments were executed on synthetic, and real images for both pixel-level and subpixel-level precision.