基于图像频域分析显著目标检测算法研究

发布时间：2018-03-16 13:00

本文选题：小波变换　切入点：频域调谐　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着科学技术的发展,手机、电脑以及其他的一些电子设备与人们的生活产生了越来越紧密的联系。人们日常生活中接触的信息也呈现指数式增长,其中视觉信息占据了绝大部分。人的视觉系统如何处理这些复杂多变的信息成为一个研究热点,与之相应的计算机视觉概念也相继提出,视觉注意机制与图像显著目标检测是其重要研究方向之一。在图像处理过程中,特征提取是十分重要的环节之一。因此,本文首先介绍了图像显著目标检测领域常用特征和提取算法。此外,对于不同的显著目标检测模型,选取不同的颜色空间会对最终的检测效果产生重要的影响,所以,本文对不同的颜色空间模型进行研究为后期模型的建立做铺垫。本文建立的第一个模型是基于小波金字塔的显著目标检测模型。该模型是受Itti生物视觉注意模型的启发,采用小波金字塔模型取代高斯金字塔模型,充分发挥图像小波变换的局部信息表达能力及多尺度空间分析能力,能够使最终的显著目标图像具有较好的轮廓信息。另外,我们将改进型的中心先验知识融入到小波金字塔模型中,进一步增强了显著目标检测效果,有利于目标分割。文章建立的第二个模型是基于图像局部分析与全局分析的显著目标检测模型。因图像的小波变换其具有上述优良的性能,能够对提取的特征图像进行局部分析,进而获得局部分析的特征显著图像。但是,由于其局部细节表达能力容易导致检测显著目标不完整以及具有复杂纹理背景信息融入,因此文章进一步融入谱残差算法对图像做出全局分析,获得全局分析的显著图像。将两种显著特征图像采用非线性融合算法进行处理,得到了最终的显著图像。从最终的实验结果分析来看,无论是直观上的显著图像还是客观上的P-R曲线,该模型的检测效果要优于其他几种算法。本文的最后一种算法是基于图像频域分析显著目标检测算法。事实上,无论是基于图像小波变换的显著目标检测模型还是基于谱残算法的显著目标检测方法,都是从图像频域角度进行分析得到的目标检测模型,文章进一步将频域调谐算法与上面的基于局部和全局分析的显著目标检测模型进行融合,进而获得了一种新的基于图像频域分析的图像显著目标检测模型。该模型在MSRA 10K以及ECSSD两个不同的数据库进行检测效果测试,结果显示,本文的模型能够较好的适应不同类型的图像,并且在图像拥有多目标,大目标以及复杂的背景的情况,检测效果都要优于其他几种算法。
[Abstract]:With the development of science and technology, mobile phones, computers and other electronic devices have become more and more closely related to people's lives. Among them, visual information accounts for the vast majority. How human visual systems deal with these complex and changeable information has become a research hotspot, and the corresponding concepts of computer vision have been put forward one after another. Visual attention mechanism and image salient target detection are one of the important research directions. In the process of image processing, feature extraction is one of the most important links. This paper first introduces the common features and extraction algorithms in the field of image salient target detection. In addition, for different salient target detection models, the selection of different color spaces will have an important impact on the final detection effect, so, In this paper, different color space models are studied to pave the way for the establishment of later models. The first model established in this paper is a significant target detection model based on wavelet pyramid. The model is inspired by the Itti biological visual attention model. The wavelet pyramid model is used to replace Gao Si's pyramid model to give full play to the ability of local information expression and multi-scale spatial analysis of image wavelet transform, which can make the final prominent target image have better contour information. We incorporate the improved central priori knowledge into the wavelet pyramid model to further enhance the significant target detection effect. The second model, which is based on local and global analysis, is a significant target detection model, because the wavelet transform of the image has the above excellent performance. The extracted feature image can be analyzed locally, and then the feature salient image can be obtained. However, because of its ability to express local details, it is easy to detect the incomplete salient target and has complex texture background information. Therefore, the paper further integrates the spectral residual algorithm to make the global analysis of the image and obtains the salient image of the global analysis. The nonlinear fusion algorithm is used to process the two salient feature images. The final salient image is obtained. From the final analysis of the experimental results, it can be seen that both the visual salient image and the objective P-R curve, The detection effect of this model is better than that of other algorithms. The last algorithm of this paper is based on image frequency domain analysis of salient target detection algorithm. In fact, Whether the significant target detection model based on image wavelet transform or the significant target detection method based on spectral disability algorithm is the target detection model obtained from the analysis of image frequency domain. Furthermore, the frequency domain tuning algorithm is fused with the above significant target detection model based on local and global analysis. Then a new image salient target detection model based on image frequency domain analysis is obtained, which is tested in two different databases, MSRA 10K and ECSSD, and the results show that, The model in this paper can adapt to different types of images, and the detection effect is better than other algorithms when the image has multiple targets, large targets and complex background.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】