针对SAR海冰图像受相干斑噪声影响严重, 提出采用相干斑抑制区域生长模型的区域MRF(SRRG-MRF)分割算法。SRRG区域模型包括构建图像的相干斑抑制区域化表达和基于区域的灰度相似性进行区域生长两个部分,其中相干斑抑制的区域化表达由相干斑抑制的双边滤波(SRBF)算法和分水岭变换构成, 该模型在相干斑噪声严重的情况下, 能够有效抑制过分割和对目标边缘准确定位, 并防止MRF分割优化陷入局部最小值, 减少误分割。将SRRG区域模型与MRF相结合, 能够大幅度减少优化搜索空间, 获得准确的分割结果。采用不同视数的SAR海冰合成图像和由RADARSAT-2及SIR-C获取的SAR海冰真实图像进行测试, 结果表明: 与已有区域MRF分割算法相比,本文算法能够有效提高分割准确性。

This paper presents a new algorithm for SAR sea ice image segmentation by combining the speckle reduction region growing(SRRG)model with region-level MRF models. Region-level MRF-based algorithms are widely used in SAR sea ice image segmentation. The over-segmentation degree of the initial areas and the localization of the target edges significantly affect the computational complexity and segmentation accuracy of region-level MRF-based algorithms. Serious over-segmentation leads to increased computational complexity, whereas accurate target edge position is beneficial to obtaining precise segmentation results. Given that existing region-level MRF-based segmentation algorithms are inadequate for determining the effects of speckle noise and the relationships between regions, the segmentation process usually needs more time, and the probability of false segmentation increases. In other words, the segmentation accuracy is usually low. Hence, this paper proposes a new segmentation algorithm based on region-level MRF models combined with a speckle reduction region growing model(SRRG-MRF). This proposed SRRG-MRF can effectively reduce the interference of the speckle noise and significantly improve segmentation accuracy by fully considering the similarity between adjacent regions.The SRRG model includes two parts: construction of an image speckle reduction regional representation and region growing based on the gray similarity of adjacent regions. For the former, speckle noise is initially suppressed using the proposed Speckle Reduction Bilateral Filter(SRBF)algorithm. The region adjacency graph is then built to obtain regional representation of the image based on watershed transform. The SRBF algorithm can thus effectively inhibit the watershed over-segmentation and achieve accurate positioning of the target edges. For the latter, the gray similarity of adjacent regions can initially describe the local characteristics of the image more accurately than the edge strength. The gray similarity penalty function of the adjacent areas is then introduced into the region MRF model based on the Gamma distribution. The region merger guideline is subsequently defined for region growing by calculating the energy difference between adjacent regions. Combining the SRRG region model with the MRF model can significantly reduce the optimization search space, prevent MRF segmentation optimization into local minimum, and reduce false segmentation to obtain accurate results.The proposed segmentation algorithm was evaluated using several synthetic SAR sea ice images corrupted with various levels of speckle noise and the real SAR sea ice images obtained by RADARSAT-2 and SIR-C, respectively. The overall segmentation accuracy and κ coefficient were used for algorithm evaluation. First, the experiment compared the number of watershed segmentation regions without filter processing and five kinds of filter processing, such as traditional Bilateral Filtering, Enhanced Lee, SRAD, and SRBF algorithm. Results showed that the SRBF algorithm is more effective in inhibiting over-segmentation and can obtain the accurate position of the target edges. The experiment then compared the proposed algorithm with the existing region-level MRF-based algorithms, namely, RMRF, IRGS, and EPR-MRF. The new segmentation algorithm substantially improved the segmentation accuracy and κ coefficient compared with the existing region-level MRF-based algorithms. The testing results demonstrated that the proposed algorithm is an effective and feasible method for SAR sea ice image segmentation. On one hand, SRRG-MRF can effectively reduce the interference of the speckle noise to inhibit watershed over-segmentation and achieve accurate positioning of the target edges. On other hand, SRRG-MRF can reduce the optimization search space, prevent MRF segmentation optimization into local minimum, and reduce false segmentation to obtain accurate results.