Some Researches On Image Fusion Application Based On Image Feature Extraction

Visual information is the primary means by which human beings get information from the nature. It is a kind of important information with subjectivity and thus difficult for computers to recognize process and implement. Image feature extraction, as an important method for computers to recognize and classify images, is the foundation of automatic image perceiving, processing and decision, and has been a hot topic in the field of computer vision for a long time. Image feature refers to those inherent characteristics or attributes including both natural features (brightness, color, texture, etc.) and artificial features (spectrum, histogram, etc.). Image feature extraction refers to the measurement of important image information which is summarized numerically.Image feature extraction is an important foundation for high-level image analysis and decision. Its performance determines the efficiency and accuracy of image processing. However, due to the complexity and diversity of countless images in real world, and different demands for image feature extraction corresponding to different purposes, there is still great room for improvement in terms of the application of image feature extraction. Further research is needed to analyze image information by image information itself, and to discover and extract image features properly and accurately, finally, to improve the quality of image processing (such as image segmentation and image fusion).Image feature extraction and its application in image fusion are discussed in this paper. Main content and innovations can be summarized as follows:1. For image feature extraction algorithms, this paper studies Independent Component Analysis (ICA) and its extensions -- Topology ICA (TICA) and Multilinear ICA (MICA). The major advantage of ICA lies in its ability to express specific characteristics of image by analyzing signals with high-order statistical properties, which is in line with the feelings of natural human visual system. As a result, features with more interests are obtained. Simulation experiments show that, ICA models have very strong abilities to describe image features, leading to better performance in the application of both image segmentation and image fusion, compared to other widely used methods.2. For color feature extraction from color images, a new color feature extraction method is proposed based on Chromaticity Distribution Feature (CDF). This method carries out Kmeans clustering with respect to color visual intensity in a*b*color space, which is closer to the feelings of human visual system, compared to in the traditional RGB space. The major advantage of CDF is to separate impact of colors according to subjective feelings. As a result, complexity of the algorithm is reduced and the extraction process is more in line with human intuitive feelings of image color characteristics.3. For color image fusion, in order to avoid artificial errors introduced by image fusion based on pixel, this study proposes image fusion methods doing image feature extraction first, and then fusing images according to those image features extracted. The study consists of two aspects: fusion algorithm based on time-domain feature extraction, and fusion algorithm based on transform domain feature extraction.(1) For feature extraction based on time domain, two methods are proposed: CDFNM, which is based on CDF feature extraction and region feature extraction, and EC fusion method, which uses evolutionary algorithm to optimize the result of feature extraction. As natural human visual system recognizes color image region by region and pixels sharing similar color are regarded as one region, the two algorithms mentioned above use regional features as input Experiments show that, compared to other traditional fusion algorithms, the two algorithms improve the quality of image fusion remarkably. In addition, artificial errors are reduced effectively. (2) With the introduction of ICA and its topology model, this paper turns to the study of fusion algorithm based on MICA feature extraction. As MICA is able to provide image feature from the point of whole picture, this paper devises a color image fusion method based on CDF and MICA, combining the advantages of these to feature description methods. Furthermore, to meet the specific demand of application, fusion rules based on region and error paradigm are suggested. The impact of significant information in characteristic region on image fusion is strengthened, and image information that is not interested in is weakened.