剪纸纹样的特征提取和识别算法研究

发布时间：2018-02-25 04:20

本文关键词： 特征提取不变矩几何特征奇异值 NMI 小波矩　出处：《广西师范大学》2010年硕士论文　论文类型：学位论文

【摘要】： 剪纸是我国历史悠久的传统民间艺术之一,随着我国动漫产业的不断发展,剪纸艺术作品将是一种很好的动漫素材。在数字媒体中去实现这种艺术形式,首先将剪纸艺术的作品转换为计算机可以储存的数字图像,进一步生成相应的作品,剪纸纹样在这些处理过程中有非常重要的作用,它决定了最终剪纸图像的艺术形态。不同剪纸纹样的精确分类与识别是剪纸纹样图像应用的基础,本文对于剪纸纹样的识别进行了研究,提出了适合于具有艺术夸张变形这种独特艺术形式的识别算法,并取得了较好的识别效果。论文的工作主要从以下几个方面展开： (1)研究剪纸图象的特点,建立了剪纸纹样库。对采集到的剪纸纹样图像进行了预处理,消除噪声,对图象进行分割,从复杂的剪纸图像中分离出单个纹样。通过分析剪纸纹样的特征,结合代数、几何、统计等方法,建立了包含有63幅训练样本和350幅待识别分类的测试样本的剪纸纹样库,这些纹样涵盖了剪纸艺术创作中的基本纹样,为后续的纹样识别工作做准备。 (2)提出一种基于不变矩与几何特征的纹样识别算法,本文使用传统的七个不变矩作为剪纸纹样的特征向量,具有平移、旋转和尺度不变性。同时又提取了纹样图像的六个几何不变特征,利用这两种不变量分别作为特征向量,采用LM算法优化BP神经网络,通过归一化后的特征向量对BP神经网络进行训练,应用训练后的神经网络作为分类器对剪纸纹样进行模式识别,试验证明该方法能够较好的识别有一定艺术变形的剪纸纹样。(3)将小波分析与其他一些特征提取方法进行了结合。小波分析具有多分辨率的特性,如果能够在小波多分辨率的基础上去提取图像的特征,那么将会提高特征向量对图像自身的表征能力。本文分别在小波分析的基础上提取了能量特征,奇异值特征,NMI特征和小波矩特征,并做了对比识别试验。小波能量方法根据小波不同分辨率下小波系数,使用能量的表示方式将低频分量和不同方向的高频分量表述出来,作为识别的特征向量。奇异值和NMI方法主要应用小波变换提取剪纸纹样图象的低频分量并进行奇异值分解或NMI提取,最后通过对特征值进行归一化和降维处理作为最终的特征向量。这种结合的方法有效的利用了小波多分辨率的特征,消除了噪声的干扰,同时又保持了奇异值和NMI特征的自身特点,且该方法计算简单,易于实现。小波矩特征具有较强的细节把握能力和抗噪声能力,通过对剪纸纹样图像提取小波矩,来获取图像的多尺度特征。利用不同特征分量的均值和标准差,来实现N类模式的特征选择。实验证明该方法能够有效地去除噪声干扰,较好的识别有一定艺术夸张变形的剪纸纹样。大量实验证明,以上方法能够很好的识别具有一定夸张变形的剪纸纹样图像,在算法的复杂度方面也能够满足计算机系统的实时要求,为下一步的剪纸图像的自动生成奠定了基础。
[Abstract]:The paper-cut is one of traditional folk art with a long history in China, with the continuous development of China's animation industry, paper-cut works will be a good cartoon material. To achieve this art form in digital media, first convert the paper-cut works for digital images can be stored on the computer, generate further work the paper-cut patterns, has a very important role in these processes, it determines the final paper-cut image of the art form. The accurate classification and identification of different patterns is based on patterns of image application, this paper does research on the identification of paper-cut patterns, put forward suitable recognition algorithm with artistic exaggeration of this unique art form, and have achieved good recognition effect.
The work of this paper is mainly carried out from the following aspects:
(1) to study the characteristics of paper-cut images, set up paper-cut patterns library. Preprocessing of paper-cut patterns image acquisition to eliminate noise of image segmentation, isolated from the patterns of complicated paper-cut image. Through the analysis of characteristics of the paper-cut patterns with algebra, geometry, statistics, set up 63 training samples and 350 testing samples for identification and classification of the paper-cut patterns libraries contain these patterns, covering the basic patterns of paper-cut art creation, prepare for pattern recognition in the future.
(2) proposed a pattern recognition algorithm based on moment invariants and geometric features, this paper uses seven traditional invariant moments as feature vector, the paper-cut patterns with translation, rotation and scale invariance. While extracting the six geometric patterns of image invariant features, the use of these two kinds of invariants are used as feature vectors, using LM algorithm to optimize BP neural network, BP neural network trained by normalized feature vectors, using the trained neural network as classifier for pattern recognition of paper-cut patterns, the experiment proves that the method can identify better have some artistic deformation patterns. (3) the wavelet analysis are combined with some of the other the feature extraction method. Wavelet analysis has the characteristics of multi-resolution, if it can be based on wavelet multi-resolution image feature extraction, it will enhance the eigenvectors of Characterization of image itself. This paper based on the wavelet analysis to extract the energy feature, singular value feature, NMI feature and wavelet moment features, and do a comparison recognition test. The wavelet energy method according to different wavelet resolution wavelet coefficients, using the energy representation of low-frequency and high-frequency components will be expressed in different directions out, as the feature vector identification. NMI method and singular value mainly used wavelet transform to extract the low-frequency components of the image of paper-cut patterns and singular value decomposition or NMI extraction, by the end of the eigenvalue normalization and dimension reduction as the final feature vector. This method combined with the effective use of wavelet multiresolution. To eliminate the noise, while maintaining the singular value and NMI features, and the method is simple, easy to implement. Wavelet moment feature has Strong ability to grasp the details and anti noise ability, through the paper-cut patterns image extraction of wavelet moments, to obtain multiscale image features. By using the mean and standard deviation of different characteristic components, to achieve N class model feature selection. The experimental results show that this method can effectively remove noise, better identification of exaggeration deformation patterns.
A large number of experiments have proved that the above methods can well identify the image of paper cut pattern with exaggerated deformation, and it can also meet the real-time requirements of computer system in terms of algorithm complexity, which lays the foundation for the next generation of automatic image generation of paper cutting.

【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2010
【分类号】：TP391.41

【引证文献】