基于局部二值模式的纹理特征研究与应用

发布时间：2018-05-18 19:47

本文选题：局部二值模式 + 特征提取　；参考：《西南交通大学》2017年硕士论文

【摘要】：纹理特征是图像的重要底层特征之一,本文对由表示图像纹理特征的纹理谱方法演变而来的LBP算法进行研究,并将改进的算法应用于图像的分类识别,目标追踪和图像分割当中。本文主要工作如下:1、研究LBP模式分类方法中的等价模式和旋转不变的等价模式,提出一种新的模式分类方式即按照0/1变换次数和二进制码值中1的数目进行分类。通过图像直方图和常用纹理库的对比试验可以看出本文提出的模式分类方法具有较高的纹理识别能力。2、用生物学中的共生概念对图像处理中的一些方法进行分析和解释,然后按照共生概念对这些图像处理方法进行分类。针对成对旋转不变的局部二值模式算法(PRICoLBP)提取方法提取的纹理特征计算复杂度高、旋转不变性较差、对较小的纹理结构特征不敏感的缺陷,提出一种改进的PRICoLBP算法。首先,改进原有算法对共生点对的选取方式,使得改进算法在保持统计更高阶纹理信息能力的同时,又增强了图像对旋转变化和光照条件变化的鲁棒性。其次,该算法融合了灰度值大小关系特征和灰度值差值幅值特征相比于原有算法只提取灰度值大小关系特征能够提取更多的纹理特征信息,从而提高了算法对较小纹理结构图像的识别能力。此外,改进算法相比于原有算法的计算维度更小。在对Brodatz,Outex,CUReT和KTH_TIPS图像纹理库的分类实验中,改进算法的识别能力相对于原有算法分别提高了 0.17%,0.24%,2.39%和2.04%。实验结果表明,改进算法在处理较小纹理结构的图像时具有较好的识别效果。针对局部二值模式特征(Local Binary Pattern,LBP)对噪声敏感、旋转不变性较差的问题,提出一种基于共生的抗噪局部二值模式纹理分类算法。首先,对LBP模式进行重新分类,对等价模式和旋转不变的等价模式进行扩展;其次,利用共生方法将原图中表示视觉微观纹理信息的LBP特征和降采样后图像中表示非视觉微观纹理信息的LBC特征进行并联,添加图像的梯度幅值信息,得到一种具有旋转稳定性和抗噪性的纹理特征表示方法。最后,在不同纹理图像库中比较本文方法和其他特征表示方法识别率的差别。实验结果表明,本文方法具有较好的旋转不变性和抗噪性。3、针对追踪过程中目标出现遮挡、目标的尺度发生变化时,STC算法容易丢失追踪目标的问题,提出一种融合LBP纹理特征的时空上下文追踪方法。首先,计算每一帧中包含目标区域的LBP纹理直方图。其次,利用卡方统计计算第一帧的LBP纹理直方图与当前帧图像内目标区域的LBP纹理直方图的相似度、相邻两帧的目标区域的LBP纹理直方图的相似度,当相似度大于设定阈值时认为目标发生遮挡,改变时空上下文更新系数和追踪目标的中心位置坐标。实验结果表明,该算法相比于原有STC算法在目标被遮挡、目标形状和尺度发生变化时均能稳健地跟踪目标,并且能够保证实时性的要求。4、针对RSF活动轮廓模型对初始轮廓敏感和分割效率较低的问题,提出了一种融合LBP纹理特征的RSF活动轮廓模型。通过引入纹理能量项,使得RSF模型对初始轮廓具有一定的鲁棒性和较快的分割速度。
[Abstract]:Texture features are one of the most important features of the image. This paper studies the LBP algorithm which is evolved from the texture spectrum method that represents the texture features of the image, and applies the improved algorithm to image classification, target tracking and image segmentation. The main work of this paper is as follows: 1, the equivalent pattern in the LBP pattern classification method is studied. A new pattern classification method is proposed, which is classified according to the number of 0/1 Transformation Times and the number of 1 in the binary code value. Through the contrast test of the image histogram and the common grain library, it can be seen that the model classification method proposed in this paper has a higher pattern recognition ability.2, and the symbiotic probability in biology is used. Some methods in image processing are analyzed and explained, and then the image processing methods are classified according to the concept of symbiosis. The texture features extracted by the local two value mode algorithm (PRICoLBP) extraction method for the pair rotation invariant are high in complexity, poor in rotation invariance and insensitive to the smaller texture features. An improved PRICoLBP algorithm is proposed. Firstly, improving the selection of the symbiotic point pairs by the original algorithm makes the improved algorithm not only keep the statistical higher order of texture information, but also enhance the robustness of the image to the change of rotation and illumination. Secondly, the algorithm combines the characteristics and gray scale of the size of the gray value. The value difference amplitude feature can extract more texture features compared to the original algorithm only extracting the relationship between the size of gray value, thus improving the recognition ability of the algorithm for the smaller texture image. In addition, the improved algorithm is smaller than the original algorithm. In the Brodatz, Outex, CUReT and KTH_TIPS image grain library, In the classification experiment, the recognition ability of the improved algorithm is improved by 0.17%, 0.24%, 2.39% and 2.04%. respectively. The results show that the improved algorithm has a better recognition effect when dealing with the smaller texture image. The local two value pattern features (Local Binary Pattern, LBP) are sensitive to noise, and the rotation invariance is poor. A kind of symbiotic denoising local two value pattern classification algorithm based on the symbiosis is proposed. First, the LBP pattern is reclassified and the equivalent mode and the equivalent pattern of rotation invariant are extended. Secondly, the LBP features of the visual microscopic texture information are expressed in the original image and the non visual micrograph is expressed in the reduced sample image by the symbiotic method. The LBC features of the texture information are parallel, and the gradient information of the image is added to obtain a texture feature representation method with rotation stability and noise resistance. Finally, the difference of recognition rate between this method and other feature representation methods is compared in different texture images. The results show that the method has good rotation. For invariance and anti noise.3, the STC algorithm is easy to lose the tracking target when the target appears to be obscured and the target scale changes. A spatio-temporal context tracking method which combines the LBP texture features is proposed. First, the LBP texture histogram of the target area is calculated in each frame. Secondly, the Chi square statistics calculation is used. The similarity between the LBP texture histogram of the first frame and the LBP texture histogram of the target area in the current frame image, the similarity degree of the LBP texture histogram of the target area adjacent to two frames. When the similarity is greater than the setting threshold, it is considered that the target is blocked, the time and space context update coefficient and the center position coordinate of the tracking target are changed. It shows that, compared with the original STC algorithm, the target can be steadily tracked when the target is blocked, the shape and scale of the target are changed, and the requirement of real-time is.4. In view of the problem that the RSF active contour model is sensitive to the initial contour and the efficiency of the segmentation is low, a RSF active contour model which combines the LBP texture features is proposed. By introducing texture energy terms, the RSF model has robustness and fast segmentation speed to the initial contour.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】