Multi-scale Local Structure Dominant Binary Pattern Learning For Image Representation

Image local feature description is a basic problem of computer vision,local feature descriptor as a low-level feature description technology,which can describle rich local details of the image,can still show strong robustness when the image is deformed,occluded or damaged,so it is widely applied in object detection and recognition and other computer vision tasks.The computer to understand of images usually adopts a "local-to-global" strategy,which is a bottom-up image description process,from Microcosmic to Mesoscopic,Macroscopic.However,the commonly used image feature extraction and description techniques are based on macroscopical whole or local region,except for LBP and GIMMRP operators,there are few local feature techniques based on microstructure description.The LBP method has the advantages of low computational complexity,rotation invariance and monotone gray scale change invariance.However,the conventional LBP method has the following defects:1)LBP has only retained the gray scale relation between pixels in the process of binarization,and large detail informations will be lost in the image;2)LBP may produce unevenly distributed histograms and low frequency patterns;3)LBP needs to be predefined for the patterns it concerns,such as "uniform pattern",which is a predefined pattern.GIMMRP has similar advantages of LBP,and has strong microstructural description ability,which greatly improves the discriminative ability of binary descriptor.However,the feature pattern coding process of the GIMMRP requires matching calculation(convolution)with the 131 templates,which results in low computational efficiency and slow speed.In addition,the important BIMP patterns obtained are the result of human selection.However,in fact,the dominant pattern and its frequency may not be different in different recognition problems and application datasets,so it is necessary to determine the most effective dominant pattern set for each data set.This paper proposes an image representation method based on image multi-scale local microstructure binary pattern extraction.By means of zero-mean microstructure pattern binarization(ZMPB),the extracted binary microstructure pattern can express all the important patterns with visual meaning that may occur in the image.Moreover,through the dominant binary pattern learning model,we can obtain the dominant feature pattern set adapted to the different data sets,which achieves excellent ability in feature robustness,discrimination and representation.Meanwhile,through dominant binary pattern learning,the dimension of feature coding can be greatly reduced and the execution speed of the algorithm can be improved.In order to verify the effectiveness of the algorithm,experiments are carried out on the ORL,YALE face data sets,MNIST handwritten digital data set and self-collected car logo data set.The experimental results show that our method has strong discriminative power and outperforms the traditional LBP and GIMMRP methods.Compared with many recent algorithms,our method also presents a certain competitive advantage.As a general description method of image recognition,it has the advantages of high efficiency,high discrimination and strong robustness.It can be widely used in target detection,object classification,text detection and recognition and other fields.