结合光谱和尺度特征的高分辨率图像边缘检测算法

高分辨率遥感图像具有高度细节化的多尺度表达能力,在有效表达地物边缘信息的同时,目标内部几何细节常以噪声的形式出现.提出将光谱相异性和小波变换相结合的边缘特征检测算法,克服了小波变换导致的边缘变形,并能够有效抑制噪声.根据光谱角原理定义归一化光谱相异性模型,并与二进小波变换结合,同时利用梯度方向余弦值对各个波段的梯度幅值加权,最后根据向量场模型计算多光谱图像的梯度幅值和梯度方向,细化后获取由细到粗的多层次边缘特征.实验结果与小波变换和传统检测算子的检测结果相比,表明该算法利用光谱相异性信息增强边缘响应强度,保证了所有尺度下获取的边缘轮廓不失真,边缘点定位准确；加权处理突出了多波段梯度主方向信息,也有效抑制了高分辨率图像上目标内部精细几何细节形成的噪声.

Edge detection of high-resolution imagery by integrating spectral and scale characteristics

The highly detailed information of objects can be provided in multi-scale by high-resolution remotely sensed imagery. When edge feature are detected in high-resolution image effectively, the internal geometric details also come to light but as noise form. In order to detect multi-scale edge feature and suppress noise, a novel method to detect the edge feature integrated spectral difference with wavelet transform was developed. Firstly, based on the theory of spectral angle, spectral difference normalized model (NSD) was defined to picture the contour of the object. Secondly, the dyadic wavelet transform was applied for each band to produce the multi-scale edge detail coefficients which actually are the gradient, and then weight the gradient magnitude of each band by using the cosine of gradient direction to enlarge the edge feature in the main gradient direction. Thirdly, combined with NSD, first fundamental form was used for detecting the gradient magnitude and orientation of multispectral images at different levels. Experiment by using QuickBird multispectral images are presented to demonstrated the representation efficiently. Compared with the results from wavelet transform and traditional edge detection operator, the proposed method can guarantee the edge without distortion, depict edge points more accurately and suppress more noise.