Key Technology Research On Military Image Segmentation

Military image segmentation is the basis and also an important part of military image processing. It’s the per-condition of the military information acquiring, military object locating, battlefield situation evaluating. To satisfy the four requirements of modern military image segmentation, which are automation, accuracy, speed and multi-target extraction, the most commonly used image segmentation algorithms:fuzzy C-means clustering algorithm, random walks algorithm, alpha-matting algorithm and snake model are studied and improved in this doctoral thesis. The main work is as follows:In the traditional fuzzy C-means clustering algorithm, the number of clustering is difficult to determine, the iteration speed is slow and the segmentation results are easy to sink into local optimal and sensitive to the initial cluster centers, so in many cases it is not fit for military image segmentation.In this thesis, a new algorithm called global best harmony search-fuzzy C-means clustering algorithm (GBHS-FCM) is proposed. The right number of clustering centers and the locations of clustering centers can be calculated automatically using the global and robust characteristics of the harmony search algorithm. Then the clustering centers and clustering numbers can be used in the FCM algorithm. A novel fuzzy clustering function is presented by combining the pixel intensity information and the spatial dependence to the neighboring pixels together, which enhanced the spatial continuity of the segmentation results. The proposed algorithm can avoid the local optimal solution of the traditional FCM algorithm, and has better performance in clustering accuracy, speed and robustness.As the traditional FCM algorithm has the shortcomings of simple image feature description, and it can be easily disturbed by complex texture with wrong segmentation, it is difficult to be applied to images with complex textures. An adaptive filtering structure tensor FCM algorithm for image segmentation is proposed. The new anisotropy filtering structure tensor is proposed to break the constraints of filtering direction and rotation of traditional filtering. Then the image edge density function for adaptively calculating anisotropy filtering proportion is combined into FCM algorithm to accurately measure image node gliding property between two adjacent nodes. Then, an adaptive filter structure tensor similarity ’measurement is defined to replace the gray level similarity measurement in the traditional FCM algorithm.To segment the images taken by satellites and romote sensing devices, a new fuzzy C-means clustering algorithm (enhanced fuzzy C-means:En-FCM) is proposed by combining image structure features to improve the ability of fuzzy C-means clustering algorithm. The input military image is mean-filtered, and the filtering image is added to the original image to form the new image for the subsequent operations. The2-D Gabor filtering function is adopted to extract texture structure feature for the new images to replace the gray level similarity measurement in the traditional FCM algorithm. Then, a new distance measure function is proposed to calculate the distance between the nodes with the clusters.To segment the images with low gray scale contrast, an improved FCM algorithm for image segmentation, combining with universal gravitation principle and local entropy theorem is proposed to overcome the shortcomings of the traditional fuzzy C-means clustering algorithm which is simple image feature description and easy distributed by complex gray influence with wrong segmentation. The image local entropy is introduced to accurately measure image node property between two adjacent nodes, and meanwhile compute the node homogeneous value. Then it is taken as the node quality, formed closely relationship using gravity algorithm which made the node gray feature and spatial position combine effectively.As the resolution of the images is becoming higher and higher and the quantities of the images are becoming larger and larger, an improved FCM combining with mean shift algorithm is proposed to improve the segmentation visual effects and efficiency of traditional FCM image segmentation algorithm. Images are segmented into many small homogeneous regions by mean shift pre-segmentation algorithm, and the homogeneous regions, instead of pixels are taken as new nodes. Then, image local entropy is adopted to describe the new nodes spatial and gray feature, after which eigenvectors of new nodes are established. Then an exponential function which can simulate well the human nonlinear visual response is used to measure the similarity between new nodes and cluster center nodes. Experimental results with both complex background and noises images show that proposed algorithm is robust to the segmentation effects and efficiency.As the traditional random walks method will lose efficiency when the background image has complicated texture features and the foreground object also has familiar texture features with the background image, the spatial features of pixels are introduce into the random walks algorithm and a novel algorithm so called spatial random walks algorithm(so called S-RW) is presented. In this new method, the new spatial feature index function is defined to measure the spatial effects between the pixels, and the new function works together with the gray level feature of the pixel to make up the edge weight function. The weight of the two effects can be adjusted by changing the freedom parameter, which makes the segmentation more accurate. The differences between pixels can be enlarged by the features defined by the marked pixel and the unmarked ones. Simulations have verified the efficiency of the present method.As the infrared images are widespreading in military, a new infrared target extracting method by alpha-matting is proposed, which considers the infrared image characteristics, so as to obtain the more subtle and accurate infrared target. Utilizing varying characteristics of successive frames of the target monitoring, and by difference computing of adjacent frames, the target prospect, background, area of uncertainty can be obtained. Adopting Quadratic Bezier Curve fitting method to establish the dynamic adjustment curve equation, the accurate transparency value of prospect was calculated, so the automatic selection of alpha value can be realized.As the multi-target images are very common in military images. A multi-target image segmentation method for fighter planes is proposed. The method combined the following three methods together:corner detection, the global best harmony search-fuzzy C-means clustering algorithm and the snake model. The corner characters of the fighter planes are identified using the corner detection, and then are clustered by the global best harmony search-fuzzy C-means clustering algorithm and the location of the planes realized. Then the improved snake model is used to extract the edges of the multi-planes.