复杂场景中高维曲线的Hough变换检测方法研究

Hough transform (HT) has been widely used as a detection approach for parametric curves and models due to its robustness and being free of heuristic information. HT usually yields large quantities of false instances when applied to real-world complex scenes. The detection performances for high dimensional curves are also to be promoted. Main research interests of this proposal are as follows. 1) The choices and rapid extraction methods of image features such as gradients, texture and visual saliency which effectively suppress the false peaks in the Hough space through weighting the votes during the HT voting process. 2) Exploration of the mothods for Hough space distribution analysis and the determination of the distribution characteristics relevant to HT detection performances, which can be used for machine learning based setting of HT algorithm parameters. 3) HT detection method based on weighting and randomization for general high dimensional curves with high time and space performances as well as high detection rates, and the exploration of the limit in the applicability of HT-based methods in high dimensional curve detection tasks. 4) Efficient hardware implementations of the proposed HT detection methods based on artificial neural networks, and case studies of applying the methods and implementations to two specific applications. The research results will enrich the relevant HT theories, and give useful clues to the understanding of the curve perception of human vision. The proposed methods can overcome some of the existing shortcomings of HT, making it applicable and promising choice for a wider range of applications.

Hough变换（HT）作为参数化曲线或模型检测方法，因其强鲁棒性和不依赖启发式信息的特点应用广泛。在复杂现实场景中，HT容易产生大量虚假结果，对高维曲线的检测性能也有待提高。本项目主要研究：1.选择适当图像特征，如梯度、纹理、视觉显著性等，对HT投票过程加权以有效抑制虚假峰值，并找到有效加权特征的快速提取方法；2.探索Hough空间分布的分析方法，获取与HT检测性能有关的空间分布特性，并据此实现HT参数的机器学习；3.基于加权和随机化等技术，寻找具有高时空性能和高检测率的一般性高维曲线HT检测方法，探索HT在高维曲线检测任务中的适用性限度；4.寻找所提HT检测方法的基于神经网络的硬件高效实现，并在两个具体问题上进行应用研究。研究成果可丰富HT相关理论，对理解人类视觉系统的相关感知过程提供有益的启发，并能克服HT的若干现有缺陷，拓宽其应用范围，提升其应用前景。

Hough变换是一种应用广泛的曲线检测方法，但在复杂现实场景中，它容易产生大量虚假结果。对参与Hough变换投票的边缘点进行加权，是一种有效的抑制虚假结果的途径。本项目研究了多种Hough变换加权方法，提出了由多尺度下边缘点处梯度方向的一致性所衡量的边缘稳定性指标，并利用该指标进行加权；提出了类LBP算子的多尺度邻域最大同号圆弧的朝向和弧度特征，并利用平均弧度和平均朝向来进行加权；提出了利用颜色显著性特征的加权方法，可利用彩色信息对边缘点进行加权。实验验证了上述加权方法的有效性。由于Hough变换算法常具有多个可调参数，因此本项目基于对Hough空间中峰值的分布的研究，提出了两种Hough变换算法参数的粒子群寻优方法。第一种方法使用有监督学习方式，根据训练集中标定的真实曲线，确定Hough空间中对应参数点附近邻域的能量比重，以之作为适应度函数完成粒子群优化；第二种方法直接以Hough空间中峰值分布的标准差作为适应度函数，并以之指导粒子群算法，对所处理的图片进行参数的自适应调整。在应用研究对若干具体应用问题进行了拓展研究，提出了高效的二值图像处理方法和骨架提取方法。项目所提出的方法，丰富了Hough变换曲线检测领域的研究，对于Hough变换在复杂场景中的应用，以及Hough变换参数调整等实际应用中的难点提供了新的思路和工具，将有助于进一步提高Hough变换在更广泛的应用领域中的适用性。