微分和形态学算子混合的图像特征检测理论与方法

In this project, we will investigate the robust and integrative feature detection theory and methods of grayscale and color images by using multi-scale, multi-directional, multi-shaped anisotropic Gaussian directional derivative (ANDD) representation as the major tool. There are three main issues for investigation. (1). We will investigate feature detection methods and algorithms by means of deep fusion of differential operators and morphological operators, so as to overcome the intrinsic defect of the traditional detectors that are not robust to impulse-type noises in images. (2). We will improve corner detection and classification methods based upon the ANDD representations and develop the blob detection and shape estimation methods using multi-scale, multi-directional, multi-shaped ANDD representations. Eventually, noise-robust corner and blob detectors are obtained. Moreover, based upon these methods, an integrative detection theory and method of image features including edges, corners, and blobs are established. (3). We will develop constant false alarm rate (CFAR) feature detection method of images corrupted by multiplicative noise by using anisotropic Gaussian directional ratio representations and establish interleaving Bayesian feature detection theory and method for images from multiple imaging systems, which provide the support in methods and algorithms to build robust scene surveillance systems using images from multiple sources.

本项目以多尺度、多方向、多形状的各向异性高斯方向导数表示为主要工具，研究灰度和彩色图像的特征一体化稳健检测理论和方法。主要的研究内容包括以下三点。(1).研究微分算子和形态学算子深度融合的图像特征检测方法和算法，解决传统基于微分算子特征检测方法对脉冲型噪声稳健性差的固有缺点。（2).完善基于各向异性方向导数表示的图像角点检测和分类方法，发展利用多尺度、多方向、多形状各向异性高斯方向导数表示的图像斑点检测和形状估计方法和算法，获得具有良好抗噪性能的图像角点和斑点检测方法。在此基础上构建灰度和彩色图像边缘、角点、斑点特征的一体化检测理论与方法。(3).发展基于各项异性高斯方向比率表示的受乘性噪声污染的图像的恒虚警率特征检测方法，建立多源图像交替Bayesian特征检测理论和方法，为发展稳健的多源图像联合场景监视系统提供方法和算法储备。

多尺度、多方向、多形状的各向异性高斯核广泛应用于图像的特征检测中，本项目以各向异性高斯核为主要工具，研究了光学图像中各向异性高斯方向导数与形态学算子的深度融合，以及SAR图像中的各向异性方向比率与形态学边缘信息的融合，提出了多种高效且稳健的光学图像边缘、斑点检测方法，以及SAR图像边缘检测和分割方法。主要研究成果包括：（1）通过对各向异性高斯方向导数和加权中值滤波器的深度融合，设计并提出了对脉冲噪声稳健的灰度边缘检测方法和彩色边缘检测方法；同时，收集50幅彩色图像，结合自动检测与人工辅助修正的方法标注每幅图像的参考边缘，构造了可以用于评价边缘检测方法性能好坏的彩色图像库；（2）结合各向异性高斯核和梯度搜索法，设计并提出了有仿射不变性的斑点测度，并在此基础上提出了对图像的仿射变换稳健的高精度斑点检测方法；（3）利用各向异性形态学方向比率和方向巴氏系数，设计并提出了两种对噪声稳健的SAR图像边缘检测方法；此外，针对SAR图像分割方法中区域合并类方法的初始分割和区域合并这两个阶段，分别提出了基于形态学边缘信息的初始分割方法和基于纹理模式关注的区域合并测度。基于本项目的研究，培养硕博研究生5名，发表12篇学术论文，其中SCI一区论文1篇，二区论文4篇。