基于视觉中心转移的视觉显著性检测方法研究

发布时间：2018-02-25 09:53

本文关键词： 视觉显著性显著图多尺度分析图像分割　出处：《西南大学》2017年硕士论文　论文类型：学位论文

【摘要】：视觉是人类认知世界获取信息的主要途径,使人能够感知复杂、变化的环境。因为人眼摄入图像的整体性和人类视觉神经系统处理信息的高度并行性,人类辨识图像并判断出其感兴趣区域是非常容易的事。随着计算机、通信和数字媒体为代表的信息技术迅速发展,视觉显著性检测已广泛应用于的遥感图像、医学图像处理、机器人视觉控制等领域。对视觉显著性检测技术进行研究与应用,使计算机具有人类视觉系统相似的信息处理能力,高效且迅速地进行图像处理,对提升图像理解系统与图像处理系统的性能,提高图像处理技术的实际应用水平都有非常重要的作用。针对现有显著性检测方法提取图像显著性目标区域的准确率以及效率较低的问题,基于人类视觉神经系统的选择性和主动性,结合图像底层的颜色对比特征、颜色分布特征、位置信息,融合多通道特征,进行多尺度分析,计算显著性特征提取出图像的显著性区域。主要工作包括以下几个方面:(1)针对传统显著性检测方法没有对图像本身先验信息加以利用的问题,提出融合背景模型和颜色特征的视觉显著性检测方法。对图像进行slic超像素分割和颜色空间转换,构造图像椭圆背景模型,在lab三个颜色通道上,分别计算椭圆内部区域的显著性特征和四个边缘背景区域的奇异性特征,线性融合不同特征通道的内部显著图和边缘背景显著图获得最终显著图。(2)针对现有检测方法提取出的显著性区域清晰程度不够,计算效率比较低的问题,提出基于视觉中心转移的视觉显著性检测方法。对图像进行slic预分割基础之上,结合图像的颜色对比特征、颜色分布特征和位置特征,提取出图像显著性区域,采用视觉转移机制模拟人眼的视觉注意中心转移过程,对图像进行多尺度分析,融合不同尺度显著图获得最终显著图。我们把上述两种方法在MSRA数据库上进行了验证实验,并和现有的检测方法进行对比。实验结果验证了我们方法的能更清晰且完整地提取出图像显著性目标区域。
[Abstract]:Vision is the main way for human beings to acquire information from the world, which enables people to perceive complex and changing environments, because of the integrity of human visual images and the high parallelism of human visual nervous systems in processing information. With the rapid development of information technology represented by computers, communications and digital media, visual salience detection has been widely used in remote sensing images, medical image processing, In the field of robot vision control, the visual salience detection technology is studied and applied to make the computer have the similar information processing ability of human visual system, and carry out image processing efficiently and quickly. To improve the performance of image understanding system and image processing system, It is very important to improve the practical application level of image processing technology. Based on the selectivity and initiative of human visual nervous system, combining the color contrast feature, color distribution feature, position information and multi-channel feature of the image bottom, the multi-scale analysis is carried out. The main work includes the following aspects: (1) aiming at the problem that the traditional salience detection method does not make use of the prior information of the image itself, A method of visual salience detection based on fusion of background model and color feature is proposed. The image is segmented by slic super-pixel and transformed into color space. The elliptical background model of image is constructed on the three color channels of lab. The salient features of the inner region of the ellipse and the singularity of the four edge background regions are calculated respectively. Linear fusion of internal salience map of different feature channels and edge background salience map to obtain final salience map. 2) aiming at the problem that the clear degree of significant region extracted by existing detection methods is not enough and the computational efficiency is low. A visual salience detection method based on the shift of visual center is proposed. On the basis of slic pre-segmentation, the salient region of the image is extracted by combining the color contrast feature, the color distribution feature and the location feature of the image. The visual shift mechanism is used to simulate the visual attention center transfer process of human eyes, and the image is analyzed by multi-scale analysis, and the final salience map is obtained by fusion of different scale salience maps. The two methods mentioned above are verified on the MSRA database. Compared with the existing detection methods, the experimental results show that our method can extract the significant target region more clearly and completely.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】