基于多尺度分析的复杂目标不变特征提取研究

发布时间：2018-03-24 15:53

本文选题：多尺度分析　切入点：人脸识别　出处：《青岛科技大学》2017年硕士论文

【摘要】：在当今社会中,人类日常生活中的目标变得越来越复杂,但无论目标变得多么复杂,目标中仍存在一些不变的特征。在对复杂目标的不变特征进行提取时,待检测特征点在不同尺度上表现形式不尽相同,因此需要在不同的尺度上对特征点进行检测,以保证检测到的特征点能够准确、有效的对目标进行描述和分析。本文的主要工作及创新性如下:(1)针对传统的特征提取方法无法在多个尺度上检测图像特征点的问题,提出了将多尺度分析与特征提取方法相结合,在不同的尺度上对原始目标图像进行特征点提取的方法。(2)在粗尺度上,选取不会产生噪声积累的形态学梯度算子对人脸图像的轮廓特征进行提取,形成一个新的人脸库—ORL轮廓人脸库,然后分别采用SIFT算法与CNN模型对ORL轮廓人脸库进行识别与匹配。使用ORL轮廓人脸库进行识别时,SIFT算法的识别率提高了1.78%,识别速度提高了34.35%;CNN模型的识别率提高了1.96%,识别速度提高了35.19%。(3)针对ORL轮廓人脸库中的人脸目标图像的纹理信息大量丢失,识别过程中容易出现差错的问题,提出了对ORL人脸库的轮廓进行提取后,进行相似度对比,将相似度较大的轮廓图像的原始图像进行整合,形成一个新的人脸库—ORL缩减人脸库,然后分别采用SIFT算法与CNN模型对ORL缩减人脸库进行识别与匹配。使用ORL缩减人脸库进行识别时,SIFT算法的识别率提高了4.78%,识别速度提高了45.80%;CNN模型的识别率提高了5.88%,识别速度提高了47.69%。
[Abstract]:In today's society, the goal in human daily life becomes more and more complex, but no matter how complex the goal becomes, there are still some invariable features in the target. The feature points to be detected are different in different scales, so it is necessary to detect the feature points on different scales to ensure the accuracy of the detected feature points. The main work and innovation of this paper are as follows: 1) aiming at the problem that traditional feature extraction methods can not detect image feature points at multiple scales, This paper proposes a method of feature point extraction from original target image at different scales, which combines multi-scale analysis with feature extraction method. The morphological gradient operator, which does not produce the accumulation of noise, is selected to extract the contour features of the face image, and a new face store-ORL contour face database is formed. Then the ORL contour face database is recognized and matched by SIFT algorithm and CNN model respectively. The recognition rate of sift algorithm is improved by 1.78 when using ORL contour face database, and the recognition rate of ORL contour model is improved by 1.96. The recognition speed is improved 35.19 / 3) the texture information of the face target image in ORL contour face database is lost a lot. It is easy to make mistakes in the recognition process. After the contour of ORL face database is extracted, the similarity degree is compared, and the original image of contour image with high similarity is integrated to form a new face database-ORL reduced face database. Then, SIFT algorithm and CNN model are used to recognize and match ORL reduced face database respectively. The recognition rate of sift algorithm is increased by 4.78 when using ORL reduced face database, and the recognition rate of ORL reduced face database is increased by 5.88, and the recognition speed is increased by 45.80%. The recognition speed increased by 47.69.
【学位授予单位】：青岛科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】