SAR Image Segmentation Algorithms Based On Fast Region Merging

Synthetic aperture radar(SAR) imaging systems have been widely applied in many fields, such as object detection and recognition, marine surveilance, cartography and nature hazard monitoring etc. SAR image segmentation is an important issue in SAR image information extraction and automatic understanding, which gives the structural information of the scene via segmenting a SAR image into several disjoint homogeneous regions. The presence of speckle noise, which are distributed randomly in SAR image plane and resulted from the inherence to coherent imaging technique of the SAR, makes degrading of the quality of the SAR image and increasing of difficulties of SAR image segmentation problem.The major studies of this dissertation are modeling of SAR image segmentation problems and its optimization. In this dissertation, several SAR image segmentation algorithms are proposed based on region merging technique. These are:1. SAR images segmentation algorithm based on edge-information-guided region merging. For the problem of the sequence of region merging in the SAR image segmentation using region merging technique, an edge-information-guided region merging technique is proposed. Firstly, utilizing of multi-direction ratio edge detector to extract the ratio edge strength map(RESM) of a SAR image, a novel thresholding method is proposed to suppress the minima in the homogeneous region of RESM, which reduces the number of regions in an initial oversegmentation produced by watershed transformation of the thresholding processed RESM. Then, a real value priority-function using areas of adjacent regions and edge information is designed to determine the sequence of region merging, which improves the precision of parameter estimation of the partition model and preserves stronger edges in the image. Finally, the proposed region merging approach is used to solve a SAR image segmentation model based on minimum description length(MDL) principle and polygonal grid representation of region boundaries. This algorithm improves the localization capability and localization precision of edges in the final segmentation results.2. SAR images segmentation algorithm based on grid code and MDL principle using region merging. A new SAR image partition model is constructed based on 8-neighbor chain grid code and MDL principle, which is fast solved using region merging. An initial partition of a SAR image is obtained combining multi-direction ratio edge detection operator with watershed transformation. Optimization of the partition model is obtained by iteratively merging adjacent region pair which leads to the most important decrease of the model’s value. Region adjacency graph(RAG) and its nearest neighbor graph(NNG) characteristic are used to speed up the proceeding of region merging. The quantitative indexes, precision(P) and recall(R), are introduced to evaluate the boundary localization capability of a segmentation methods. The experiments show that the method has higher boundary localization capability and lower computational complexity.3. SAR images segmentation algorithm based on G0 distribution and chain code grid. An adaptive weighted SAR images segmentation model is proposed based on MDL principle and to alleviate the influence of complexity of SAR images’ scene on segmentation results. In the model, G0 distribution is used for describing SAR image data and chain code grid is utilized for coding boundaries of regions in the SAR image. An adaptive estimation method of the weight of the segmentation model is proposed using the SAR image data. The model is efficiently minimized by iteratively merging the adjacent regions decreasing the model’s value most importantly in initial segmentation. The experimental results show that, this method effectively alleviates the degree of over-segmentation in texture fields.4. SAR image segmentation algorithm using hierarchical region merging with edge penalty. A new SAR image segmentation model with edge penalty is constructed, which uses oriented edge strength information and is minimized by a proposed hierarchical region merging algorithm. The edge strength information is extracted by using multi-direction ratio edge detector, on which a high quality initial over-segmentation is obtained using watershed transformation. In order to extract the directions of boundaries of regions, polygons are used to approximate them, and a penalty term whose power is in inverse proportion to edge strength is obtained byincorporating oriented edge strength map(OESM) into the term. A hierarchical region merging algorithm driven by image features is obtained through graduated increased the strength of edge penalty. In order to accelerate the region merging, the region adjacency graph(RAG) is used to represent image segmentation results. The experimental results show that, with respect to other methods, this method has advantages in performance and efficiency, and obtains better segmentation results.5. SAR images segmentation algorithm via region merging with relative common boundary length penalty. A region-merging-based method is proposed for fast segmentation of SAR images. It combines the existing fast initial partition by applying the watershed transform to the thresholded ratio edge strength map(RESM) with fast region merging by using a new merging cost with relative common boundary length penalty(RCBLP) and the nearest neighbor graph(NNG) for fast search minimal edge in a graph. A new statistical similarity measure, which is scale invariant and approximate constant false alarm rate with respect to region sizes, is proposed and combined with a RCBLP term to form a new merging cost. The region merging process is fast implemented by the means of the region adjacency graph(RAG) and NNG. The quantitative indexes precision(P) and recall(R) are used to assess the performance of edge localization and the segment covering criterion is used to assess the performance of detection of regions. Experiments to synthetic and real SAR images are reported. The results show that the proposed method is fast and attains higher-quality segmentation results than the two recent state-of-the-art methods.