生物视觉启发下的边缘检测方法及其应用研究

发布时间：2018-04-12 07:53

本文选题：生物视觉 + 格式塔认知规则　；参考：《华中科技大学》2016年博士论文

【摘要】：边缘检测是图像处理中的基础问题,也是经典的技术难题之一,其结果对于特征提取和描述、目标检测和识别、图像理解等后续处理和应用有非常重要的影响,因此开展相关问题的研究不论在理论方法上、还是在应用层面都有重要的意义。由于成像条件的多样性以及视觉场景的复杂性,比如噪声、畸变带来的图像模糊以及复杂背景对目标的边缘的干扰,虽然已经提出了大量的边缘检测算法,但无论是边缘检测的效果,还是基于边缘的目标识别等应用,现在的机器视觉系统的性能和适应性上仍然与人类视觉系统存在很大的差距。另一方面,近年来对生物视觉的研究取得了很大的进展,研究者通过神经生理学和心理学方面的大量研究,对视皮层的结构功能、视觉信息的处理通路和视觉认知的规律等方面取得很多研究成果,这些成果为边缘检测算法的设计提供了新的理论指导。论文以生物视觉理论为指导,通过模拟人类视觉系统的感知特性和认知方法,从边缘检测、轮廓检测、轮廓编组等方面对边缘检测的相关问题进行了深入系统研究,并将其成功应用于自动目标识别。论文的主要研究内容如下：首先,针对传统边缘检测算法中模板形状和大小难以确定的问题,在不同神经元细胞感受野特性的启发下,提出了一种基于感受野特性的边缘检测算法。该算法通过模拟侧膝体、结合非经典感受野特性的简单细胞以及复杂细胞的响应特性,实现边缘检测。其中,针对噪声对边缘检测结果的影响,通过对简单细胞感受野形成机制的研究,结合非经典感受野特性对感受野合并模型进行改进来模拟简单细胞感受野响应,相比于传统的简单细胞响应模型,该模型不仅可以更好地模拟简单细胞的生理结构,而且综合考虑了非经典感受野的易化和抑制作用,从而可以更好地增强边缘抑制噪声。经过实验验证,该算法对于噪声和背景干扰有很好的鲁棒性。其次,针对轮廓检测算法中重复计算和结构信息利用不足的问题,提出了一种基于多层视觉线索的轮廓检测算法。不同于在像素级进行检测的传统方法,该方法首先基于超像素的方法生成候选轮廓集,结合生物视觉层级处理流程和相关理论,通过提取多层视觉线索评价候选轮廓集来实现轮廓检测。实验结果表明,该算法可以有效地检测图像中目标的显著轮廓。然后,针对传统轮廓检测算法无法区分多目标图像中不同目标轮廓的问题,在格式塔认知规则的指导下,以轮廓检测为基础,对多目标轮廓编组问题进行了进一步的研究,以轮廓检测的结果为编组基元,提出了一种基于格式塔编组约束和谱聚类的多目标轮廓编组方法,利用编组约束构造的亲和矩阵自动估计图像中目标的数目,引入谱聚类的方法实现对不同目标轮廓的编组,将属于不同目标的轮廓分别标记出来。通过实验验证,该方法可以有效的估计目标数目,并实现对图像中多个目标轮廓的编组。最后,针对传统显著性目标区选择方法对目标边缘和轮廓信息利用不足的问题,提出了一种基于边缘分布显著图的候选目标区选取方法。在利用自动目标识别技术进行导航制导的应用中,为了使识别算法发挥最大的功效,除了对识别算法的性能有较高的要求以外,还必须考虑选择合适的区域,这样才能使高性能的识别方法发挥最大的功效。从这一角度出发,在复杂细胞汇聚作用对不变特征提取的启发下,构造了一种反映边缘方向分布的特征,并且利用该特征提出了一种基于边缘分布显著图的候选目标区选取方法。通过在多种影像和应用上的测试,验证了提出算法的有效性和实用价值。
[Abstract]:Edge detection is a fundamental problem in image processing, one of the technical problems is classic, the results for feature extraction and description, target detection and recognition, has a very important impact on image understanding and subsequent processing and application, so the study on the related problems in theory, still have important significance in the application because the imaging condition level. The diversity and complexity of the visual scene, such as noise, interference caused by fuzzy image distortion and complex background edge on the target, although the edge detection algorithm has been proposed, but whether the effect of edge detection, or application of target recognition based on the edge of the machine vision system now the performance and adaptability of the human visual system and still have a large gap. On the other hand, in recent years the study of biological vision has made great progress. Through a large number of researchers studying neurophysiology and psychology, the structure and function of visual cortex, made a lot of research on visual information processing pathways and visual cognition rules. These results provide a new theoretical guidance for the design of edge detection algorithm. This paper uses biological vision theory as the guidance, through the perception and cognitive method from the simulation of the human visual system, edge detection, contour detection, contour grouping etc. the problems related to edge detection is studied deeply, and applied successfully to automatic target recognition. The main research contents of this paper are as follows: firstly, according to the template of traditional edge detection algorithm is difficult to determine the shape and size of the problem, inspiration receptive field properties in different neurons, presents a receptive field properties of the edge detection algorithm based on the algorithm. Through the simulation of the knee Body, simple cells with non classical receptive field characteristics and response characteristics of complex cell edge detection. Among them, the influence of noise on edge detection, through the research on the formation mechanism of the wild simple cells, combined with non classical feelings improved to simulate the simple cell receptive field response of receptive field with wild model simple cell response, compared to the traditional model, the physiological structure of the model can not only simulate the simple cells, and considering the non classical receptive field facilitation and inhibition, which can better enhance the edge noise. After experimental verification, the algorithm has good robustness to noise and background interference secondly, aiming at the shortage problem and structure information using the repeated calculation of contour detection algorithm, proposes a multi-layer contour detection algorithm based on visual cues. The traditional method for testing the same as in the pixel level, the first super pixel method to generate candidate sets based on contour, combined with biological processes and related theory of visual hierarchy, to achieve contour detection by extracting the evaluation of multi visual cue candidate contour set. The experimental results indicate that significant contours of this algorithm can effectively detect the target image. Then, according to the different traditional contour contour detection algorithm can not distinguish between multiple targets in the image, the Gestalt cognitive rules under the guidance of the contour detection based on multi target contour coding group was further studied, the contour detection results for grouping primitives, presents a Gestalt grouping constraints based on spectral clustering and multi target contour grouping method, automatic estimation of the number of objects in an image using the affinity matrix constructed by grouping constraints, spectral clustering. Method of grouping of different target contour, belong to different target contour respectively marked. Experiments show that this method can effectively estimate the target number, and grouping of multiple target contour in the image. Finally, aiming at the significant target area selection method of target edge and contour information use the problem, put forward a method of selecting candidate target region saliency map based on edge distribution. Application of navigation guidance in the use of automatic target recognition technology, in order to make the recognition algorithm for maximum effect, in addition to the performance of the recognition algorithm has higher requirements, must also be considered to choose the right area, so as to make the the recognition method of high performance play the biggest effect. From this point of view, in the role of stimulating the complex cell aggregation feature extraction, construct a reflecting edge side According to the characteristics of distribution, and using this feature, a candidate target area selection method based on edge distribution saliency map is proposed. The effectiveness and practicality of the proposed algorithm is verified by testing on various images and applications.

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】