低质量指纹图像增强与特征提取技术.pdf

国防科学技术大学 硕士学位论文 低质量指纹图像增强与特征提取技术 姓名：周海徽 申请学位级别：硕士 专业：电子与通信工程 指导教师：李吉成 2011-03 国防科学技术大学研究生院工程硕士学位论文 第 i 页 摘 要 目前，指纹识别技术已广泛应用于门禁控制、信息保密、远程认证等领域 这些领域采集到的指纹图像清晰、完整且对比度高然而，针对作案现场提取的 罪犯指纹，指纹通常是不完整的，且纹线的对比度也很低对于这种低质量的指 纹图像，用传统的指纹识别技术处理，很难得到理想的效果 本文以低质量指纹图像为研究对象，对自动指纹识别系统中的关键技术进行 了深入研究，主要包括低质量指纹图像的分割、增强和特征提取等技术 主要工作和创新如下： （1）针对传统的基于块和像素点的指纹图像分割算法对低质量指纹图像分割 效果不佳，在分割过程中容易产生错误分割导致白块效应和有效区域不能与背景 完全分离的缺点，提出了基于Canny算子与形态滤波相结合的低质量指纹分割方 法 该方法利用Canny算子求取边缘来完成前景模板的粗分割， 然后运用基于数学 形态学的形态滤波技术，通过选择合适的结构元素对指纹有效区域进行边缘修正 和噪声去除，从而准确地将有效区域和背景区分离，且不会造成指纹有效区域脊 线信息的丢失。

（2）研究了基于方向滤波和传统Gabor滤波的增强技术，实验结果表明，对 于低质量指纹图像，传统Gabor滤波比方向滤波有更好的增强效果针对传统 Gabor滤波中纹线方向参数取值不够精确、纹线频率参数求取方法效率不高的缺 点，本文对分块方向图通过不同尺度分块的方向图进行判别改进，获得了相对精 确的方向参数；采用了一种新的求取纹线频率的方法，用平均频率代替传统的纹 线频率参数，提高了其求取的效率，并采用了一个可以根据局部纹路方向和频率 来调整滤波窗口大小和方向的算法对指纹图像进行增强仿真实验表明，改进的 Gabor滤波算法在增强效果和实时性方面优于传统的Gabor滤波 （3）采用自适应二值化技术，对增强后的指纹图像进行局部自适应阈值的二 值化处理对于处理后的二值图像，针对自适应阈值求取过大容易在谷线上产生 粘连和阈值求取过小容易在脊线上产生气泡的缺点，提出了根据指纹块方向去除 粘连和根据白点邻域填充气泡的去噪算法对于去噪后的二值图像，采用基于改 进的OPTA指纹细化算法， 对细化后的单像素脊线指纹图像， 选择端点和分叉点作 为低质量指纹的细节特征，采用模板匹配法进行特征提取，并对提取的细节特征 作去伪处理。

实验表明，去噪处理后的二值图像更为清晰，纹线也更加平滑 主题词：指纹识别，指纹分割，指纹增强，特征提取 国防科学技术大学研究生院工程硕士学位论文 第 ii 页 ABSTRACT So far, the existent fingerprint recognition researches both at home and abroad are always concentrated to common applications, such as the gate controlling, the information security, remote confirmation, which demand for the clearness, integrity, and high contrast of fingerprint images. However, the criminal’s fingerprints withdrew from the spot are not sufficient, and even worse, the contrast of the textures is also low. Therefore, it will be difficult to achieve desired results by using traditional techniques to process such poor quality fingerprint images. This paper is mainly researched on poor fingerprint images, and has made further researches on the key technologies of the fingerprint auto-recognition system, such as the segmentation, enhancement, and feature extraction techniques of these images. The main work and novelty of this paper is: (1) Since the traditional fingerprint’s segmentation algorithm based on block and point, is prone to make error segmentation so as to get white-block effect and can’t make effective segmentation between the effective area and background. Then this paper proposes a new segmentation algorithm for the low-quality fingerprint’s efficient area, which combines edge extraction with morphologic filtering techniques. The algorithm uses Canny operator to get the edge, so as to achieve the coarse segmentation of the foreground model. Then a morphologic method is employed. An appropriate structural element is selected to correct the edge in the effective area and reduce noise. Then it will segment the background and foreground not only effectively but also exactly. By experiment with fingerprint, this algorithm can reduce the noise of the background, while avoiding the losing of ridge line’s information in the effective areas. (2) Make researches on two traditional image enhancement techniques: directional filter and traditional Gabor filter. By experiment, it shows that traditional Gabor filter can achieve better enhancement effect than directional filter when processing low-quality fingerprint images. Since the traditional Gabor filter has such drawbacks as low preciseness of the direction parameters of the texture lines, and the low efficiency of the resolution of the frequency parameter, the blocked directional image is distinguished by using directional images with various scales to obtain more precise direction parameter. A novel method is employed to get the mean frequency of the ridge lines. Unlike the traditional Gabor filter using the texture line frequency parameter, mean frequency is used instead, and the efficiency of the resolving is improved greatly. Lastly, the filtering window of the traditional Gabor filter is improved. A new filtering window , which can adjust its size and direction adaptively according to the direction of local texture and frequency, is used in the enhancement of the fingerprint images. When enhancing images based on segmentation, it shows that the improved method has more 国防科学技术大学研究生院工程硕士学位论文 第 iii 页 privilege over former ones both on enhancement effect and time costing. (3) Binarize the enhanced fingerprint image using binary techniques based on local adaptive threshold. For the attended binary images, on one hand, if the threshold is large excessively, it will be prone to get sticks on the valley line; on the other hand, if the threshold is small excessively, it is prone to get bobs on the ridge line. Then a denoising algorithm is proposed, which eliminates the sticks according to the block direction, and fills the bobs according to the neighborhood. Utilize an improved fingerprint image thinning algorithm. After thinning, texture lines become smooth and clear, and at the same time hold the texture construction of the low-quality fingerprint images. For those thinned images, select the end minutiae and。