Research On Image Line Feature Detection And Description

With the rapid development of computer hardware and software technology, thecomputer vision technology has been widely used in daily entertainment, industrialproduction and military application. However, these applications only work under somespecific environments, and there is no once-and-for-all solution for the majority ofcomputer vision problems. One of the reasons is that it is difficult to develop a genericimage feature descriptor. So we still need to construct more robust and effective imagefeatures.As the most common low-level image features, image edge contains a wealth ofinformation. There exist rich edges in the image under the artificial environment. Howto extract and describe these edges? How to use this information in the computer visiontasks such as image registration, pattern recognition, image segmentation andsuper-resolution? Based on this consideration, we focus on the image line featuredetection and description in the thesis. The main work and contributions are as follows:1. We propose an edge detection algorithm based on Two-Step Optimal edgegrowing strategy. To ensure the continuity of the edge detection under low-contrastconditions, we first apply an adaptive threshold to the image gradient magnitudeaccording to Weber–Fechner Law. Then we use a Two-Step Optimal edge growingstrategy to extract continuous, stable and single-pixel width image edge. Furthermore,we propose a fast parameter-free line segment splitting algorithm, which can fit thecurve segment by several line segments. We use these line segments to build image edgeline descriptor in the following step.2. We propose two kinds of descriptors based on the edge line segment. One is amixed gradient line feature descriptor, another is a descriptor based on a new binarypattern schemes. We use the edge gradient amplitude distribution and edge segment’slength to determine the edge’s scale. Besides, we implement a secondary adjustments tothe descriptor based on the spatial distance and the edge strength. We finally obtain theimage edge feature that is invariant to rotation, scale, translation, brightness andcontrast. 3. We study the matching performance of the proposed line descriptor in manydifferent cases. Besides, we construct the global image feature based on the local linefeature. We also analyze the characteristics of our proposed global feature for thebuilding style classification.The proposed edge detection algorithm in this thesis can generate a fast and stableedge without glitches and bifurcations. Moreover, it can output a structural edge result,which is very critical for the post-processing. The proposed line descriptor in this thesisperforms well both in describing the local and global image edge, which is validated inthe feature matching and image classification tasks.