Research On Multi-spectral Image Fusion And Processing Technique

Recently, with the fast development of multi-spectral imaging instruments, multi-spectral image has been utilized in many fields such as military, meteorology, medicine and remote sensing and it plays more and more important role in life and manufacture. Accompanying with the increment of sensor types, the quantity and the diversity of image data grow up rapidly. How to process multi-spectral image and make use of it efficiently is a valuable topic which needs to be studied.This thesis presents some techniques on potential object area detection, image thresholding, image fusion on pixel and feature level, which serve for object detection and recognition in middle wave infrared image, long wave infrared image and visible image. At last, all the research achievement listed above has been integrated together with some other techniques to form an integral system for multi-spectral image processing. The main work and innovation are listed as follows.In object detection and recognition, potential object area detection is an important step, which could reduce the difficulty of detection and recognition and cut down the data quantity. Remote sensing classification is an efficient tool for potential object area detection. In this thesis, the improved run-length features are applied to remote sensing classification combined with neural network. In feature selection, rough set and the variance within-class and inter-class are utilized. The experiment demonstrates the efficiency of our method.Image thresholding is a simple and effective way for image segmentation, which could be widely used in feature extraction and object detection and recognition. In order to utilized the gray value and spatial distribution information of pixels meanwhile, gray level spatial correlation histogram (GLSC histogram) is proposed in this thesis, and the human visual nonlinearity characteristics is embedded. Here, based on GLSC histogram, two thresholding approaches using information entropy and type-2 fuzzy set are proposed. Compared with traditional ways, our methods could achieve balance in segmentation performance and time consumption.The contrast between object and background has significant effect on object detection and recognition. In order to enhance the contrast, an image fusion method on pixel level using nonsubsampled contourlet transform (NSCT) is proposed. The fusion rule for high-frequency and low-frequency component is formulated to be beneficial for object detection and recognition.Image fusion on feature level is a high level fusion approach. The key to feature fusion is how to fix the redundancy and the complementary between multi-spectral features. Feature fusion could improve object detection and recognition accuracy by composing features with high complementary and low redundancy. Based on locality-constrained linear coding (LLC), a new image fusion method on feature level is suggested in this thesis. The redundancy and the complementary could be analyzed by LLC and feature fusion is executed by max-pooling. In specific object detection and recognition, HOG feature is extracted and modular SVM is employed as classifier.In order to integrate all the techniques listed above and some other research achievement, a multi-layer integrated system for multi-spectral image processing is proposed in this thesis. The system includes functions as image registration, feature extraction, multi-spectral image fusion, object detection and recognition and so on.