Gaussian反对称小波与SAR影像目标特征提取

Gaussian Antisymmetric Wavelets Built for Extracting the Objects and Features in SAR Image

由于遥感影像中目标尺度的不确定性 ,使得在目标特征提取和检测前难以事先确定最佳尺度以适应影像空间不同尺度目标的识别与检测 ,给遥感影像自动识别与全自动数字摄影测量带来了极大困难。为了解决遥感影像中不同尺度目标的探测问题 ,提出基于一般高斯核构造一类Gaussian反对称小波并给出相应二维小波变换的快速算法 ,弥补了MallatGaussian小波 (σ =1)在解决遥感影像空间不同尺度目标检测中的不足。对于影像中不同尺度的目标特征探测 ,通过在Gaussian核函数中选择适当的参数σ值 ,得到相应的反对称小波空间滤波器响应系数 ,用于解决不同目标的检测问题。分别给出了有关Gaussian反对称小波的 5组空间滤波器响应系数 ,对于研究遥感影像目标兴趣算子和多源遥感影像特征检测尤为重要。由于SAR影像中“speckles”是一乘性噪声 ,文中先对原SAR影像进行对数变换得到一同态影像。基于小波变换的软域值斑点噪声抑制既在一定程度上抑制斑点噪声又保持影像细节信息。通过对两SAR影像检测实验 ,表明给出的Gaussian反对称小波类在遥感影像特征检测中效果极佳。遥感影像中同时存在阶跃型和屋脊型边缘 ,可以采用反对称小波的极大模或过零点检测提取影像边缘特征 ,但得到的结果存在局部位置偏差。

It is well-known that, because the object scales in remote sensing images change over wide and unpredictable ranges, a problem in selecting adaptive scale filter is existed for extracting different scale objects in remote sensing images. Aiming at object scales in remote sensing image change uncertain, we introduce one class of Gaussian antisymmetric wavelets based on Gaussian kernel, which extends to Mallat Gaussian wavelet (σ=1). Coefficients of spatial filter related to the class of Gaussian antisymmetric wavelets given in the paper are derived adaptively by selecting appropriate parameter σ values for special scale object extraction in SAR. Five group coefficients of spatial filter related to the antisymmetric wavelets have been given in this paper. It is important that ones develop interesting operators for the object recognition of SAR images and investigate approaches for feature detection in multi-resource remote sensing images. Because `speckle' in SAR images is a multiplicative noise, we performed firstly logarithm transform over the two SAR images in preprocessing. Then the features in the logarithm images may be detected in the wavelet transform. It is shown by our experiments in two SAR images that the class of Gaussian antisymmetric wavelets is very efficient for feature extraction in remote sensing images, in which object scales change over wide ranges. Because there exist both step and roof edges in remote sensing images, maximum modulus, or zero-crossings of antisymmetric wavelet transforms can be used for edge feature detection, but the results detected exist local position discrepancy. The conclusion is important to explore new edge detectors in remote sensing images and new technology related to all automatic digital photogrammetry in future.