Multi-scale Feature Detection: Methodology And Application

Generally, a signal and information processing system will consist of 3 modules, i.e. signal preprocessing, feature detection\quantification, and pattern recognition, among which the second module aims to detect related feature from preprocessed signal, and to quantify them into some numeric descriptors for latter pattern recognition. Therefore, feature detection\quantification is the key module, whose sensitivity, robustness, and precision will decide the whole system performance. Due to the diversity and complexity of nature, feature detection\quantification in a multi-scale style isan important technical problem, which has been wildly studied and used in recent years throughout many fields, e.g. biology, astronomy, geography, and etc.There are two fundamental problems lies in multi-scale feature detection:By which operator could we introduce a multi-scale representation? How to fuse all the information in different scales for feature detection? The first problem has been already solved by linear scale-space theory and its extensions. However, there is no unified theory on the second problem, whose related research has been scattered in all kinds of literatures across different fields.The main work and innovation points of this dissertation can be divided into the following two parts, i.e. research and applications.In the methodology part, after a comprehensive literature review, we present an algorithm framework of multi-scale feature detection and fusion for different applications. The framework is composed of three stages, i.e. mono-scale feature detection, scale-space deep structure construction, deep-structure cost function design for final feature decision. However, there is no universal theoretic guide for the design of each stage, where the specific application objective and background knowledge should be taken into account.In the application part, the above algorithm framework of multi-scale feature detection and fusion is applied to both protein mass spectrometry (MS) analysis and film defect detection so as to verify its effectiveness and to solve some common problems and challenges in the two application fields. Our main contributions and innovations are as follows. In protein MS analysis, the multi-scale framework is adopted for peak detection to resist spectrum variations as the main challenge for current MS analysis systems. Specifically, we first introduce peak tree as a scale-space deep structure to represent the relationship among peak features on different scales. After that, peak detection is converted into a problem of peak tree decomposition, whose cost function is based on both local peak intensity and global common peak information so as to resist spectrum variations. Based on the two points, we get a novel closed-loop peak detection framework, which finds the optimal detection resultvia iterative refinements. Finally, we present a feature selection method by stochastic ant colony optimization (ACO) with multivariate-statistical heuristics and deterministiclocal optimizations. Our final MS system is based on peak-tree peak detection and ACO feature selection, whose benefits on both peak detection and MS disease analysis are proved on both virtual spectrum and real SELDI data.In film defect detection and reparation, scratch is a common type of film detects which is hard to detect. Our goal is to find an automatic roust detection method which can adapt to different scratch widths, locate the scratch region precisely, and differentiate scratch from straight edges. To this end, we first design a complex Ridgelet filter and its related multiscale multiscale detection algorithm for line feature, which can precisely locate scratches of different widths and recognize straight edges at the same time. Then, we present a general uni-modal threshold selection method based on Weibull distribution model to make the whole system automatic. Finally, we present a scratch verification method based on neighborhood color distance to improve the robustness of the system. Benefits from the above algorithms are proved by both stimulated and real data.