噪声背景下线状目标的编组提取方法研究

The extraction of curve structures, also referred to as curves, is a fundamental image operation, which has many applications in computer vision and pattern recognition. Traditional methods for curve structure extraction can be divided into two categories. One is based on edge detection, and the other is based on segmentation. However, great challenges are raised when extracting curve structures with different intensity, different width, and arbitrary shapes from noisy background, by solely using edge information or region information. In this project, a novel method based on perceptual grouping will be studied, where the problem of curve structure extraction is formulated as a problem of contour grouping by combining edge information and region information. First, a saliency model for curve structures is defined to provide a unified solution to produce the linear saliency of curve structures regardless of their intensity or width. Then, a class-oriented learning method for edge detection is exploited to detect faint edges from noisy background. Finally,a grouping model incorporating both edge information and region information(i.e., linear saliency) is studied, and graph techniques and optimization algorithms are employed to extract curve structures with arbitrary shapes. The research of the project will form a unified approach to solve the problem of extracting curve structures with different intensity, different width, and arbitrary shapes.

线状目标提取是一种基本的图像操作,它在计算机视觉与模式识别中具有重要的应用。传统的线状目标提取方法可分为两类，一类基于边缘检测，另一类基于区域分割。然而，在噪声背景下提取不同亮度、不同宽度、形状不规则的线状目标时，单独利用边缘信息或区域信息的方法遇到了严峻挑战。本项目计划研究一种新的基于感知编组的线状目标提取方法，将线状目标提取问题转化为结合边缘信息和区域信息的轮廓编组问题。主要研究内容包括：1）设计针对线状目标的显著度计算模型和算法，实现对不同亮度、不同宽度的线状目标在线性显著度上的统一度量；2）研究基于分类学习的边缘检测算法，实现对噪声背景下线状目标的边缘检测；3）推导结合边缘信息和区域信息(线性显著度)的编组模型，利用图论技术及最优化计算，实现对不规则线状目标的完整提取。该项目的研究成果将形成一种统一的方法，解决对不同亮度、不同宽度、形状不规则的线状目标的提取问题。

噪声背景通常给线状目标检测与识别带来了困难。本项目针对传统的边缘检测、阈值分割在检测现状目标时存在不可靠性的缺点，研究了一种新的基于感知编组的线状目标提取方法。主要研究内容分为五个部分：（1）对不规则形状路面阴影的自动消除方法。在基于亮度高程划分的亮度补偿模型下，利用阴影区、半影区和无阴影区的亮度曲线变化特点，提出了自动查找最优参考区的阴影消除方法；（2）研究了基于深度卷积神经网络的线状目标显著度计算方法。利用深度全卷积网络构建线状目标的分割模型，利用仿真的大数据量样本学习网络参数，然后利用真实的裂缝影像对网络模型进行精化（fine-tune）；（3）利用边缘信息和亮度区间信息，构建了基于路径投票的线状目标编组提取方法，为了减轻传统的基于最短路径搜索法对亮度的过分依赖，提出了基于图像分割的路径投票算法；（4）提出了一种区域颜色共生特征描述子，能有效描述图像中的颜色搭配，实现对彩色图像的鉴别性编码和准确分类；（5）提出了一种基于深度学习的特征选择方法，利用深度置信网的特征抽象功能，将重构误差小的特征作为分类特征，提高了遥感影像场景分类的精度。实验结果表明，所提的影像消除方法可以在保持线状目标的同时去除阴影；深度卷积网络能有效用于线状目标的显著度计算；所提的基于分割的路径投票算法，实现了对不同亮度、不同形状的线状目标的增强与提取；所提颜色共生描述子能有效描述颜色的搭配模式，取得比现有颜色描述算子更高的分类精度；基于深度学习的特征选择方法，能得到具有类别鉴别性的特征，提高场景的分类精度。