摘要
针对原始的局部相位量化(Local Phase Quantization,LPQ)算法对具有模糊不变性的相位特征描述不准确、缺少对图像重要细节信息描述的缺点,提出了一种结合高斯拉普拉斯(Laplace of Gaussian,LoG)边缘检测和增强局部相位量化(Enhanced Local Phase Quantization,ELPQ)的模糊图像识别算法,记为MrELPQ&MsLoG(Multi-resolution ELPQ and Multi-scale LoG)。首先,在频域中,将图像进行短时傅里叶变换后得到的实部与虚部进行正负量化和幅值量化,得到互补的符号特征ELPQ_S和幅值特征ELPQ_M;其次,在空间域中,利用多尺度高斯拉普拉斯与图像进行卷积得到图像空间域的边缘特征;最后,将频域上的符号特征ELPQ_S和幅值特征ELPQ_M与空间域上的边缘特征结合,生成最终的特征直方图,采用SVM进行识别。在有模糊干扰的Brodatz和KTH-TIPS纹理库中,文中提出的ELPQ算法相比原始的LPQ算法有较大的性能提升,且空间域和频域结合的MrELPQ&MsLoG算法能进一步提高算法的识别性能;在具有模糊的AR、Extend Yale B人脸库和实际拍摄的铁路扣件库中,将MrELPQ&MsLoG算法与目前模糊鲁棒性较好的算法进行对比发现,MrELPQ&MsLoG算法保持着较高的识别率。实验结果表明,MrELPQ&MsLoG算法对模糊具有较强的鲁棒性,且特征提取时间较短,具有实时性。
As ablur insensitive texture descriptor,Local phase quantization(LPQ)algorithm describes phase features with blurred invariance in accurately.Besides,it lacks in describing important details of images.In order to solve the issues,an enhanced local phase quantization(ELPQ)combined with Laplace of Gaussian(LoG)edge detection is proposed in this paper,named MrELPQ&MsLoG(Multi-resolution ELPQ and Multi-scale LoG).Firstly,the real and imaginary parts obtained by performing the short-term Fourier transform on the image arepositive and negative quantification and amplitude quantization,complementary symbol feature ELPQ_S and amplitude feature ELPQ_M are obtained.Secondly,the edge features in spatial domain are obtained by convolving images with multi-scale Laplace of Gaussian filters.Finally,the symbol feature ELPQ_S and the amplitude feature ELPQ_M in the frequency domain are combined with edge features on the spatial domain.The recognition result is calculated through SVM.On the Brodatz and KTH-TIPS texture data bases with blurred interference,the ELPQ algorithm has a great improvement over the original LPQ algorithm.Moreover,the MrELPQ&MsLoG algorithm can further improve the recognition rate of the algorithm.On the AR,Extend Yale Band railway fastener data bases with blurred interference,compared with the current algorithm which has robustness to the blur,the MrELPQ&MsLoG algorithm always maintains a high recognition rate.The experimental results show that the MrELPQ&MsLoG algorithm is robust to blur and has less time for feature extraction.
作者
陈晓文
刘光帅
刘望华
李旭瑞
CHEN Xiao-wen;LIU Guang-shuai;LIU Wang-hua;LI Xu-rui(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China)
出处
《计算机科学》
CSCD
北大核心
2020年第12期197-204,共8页
Computer Science
基金
国家自然科学基金资助项目(51275431)
四川省科技支撑计划项目(2015GZ0200)。
关键词
局部相位量化
边缘特征
模糊鲁棒性
频域
空间域
Local phase quantization
Edge feature
Blurred robustness
Frequency domain
Spatial domain