Binaryzation Express Of Image Micorstructure And Object Detection And Recognition Application

The image feature representation is a very important basis work for object detection and recognition.At present,the commonly used image features are divided into two ways: global and local.The local feature show strong distinctiveness,stability and localization is characterized by fine description ability for the local area of the image,and show strong distinctiveness,stability and localization.Local features can solve a lot of difficult computer vision problems,and has become the mainstream methods for object detection and recognition.It is not difficult to find that the real images can be understood from three levels: pixel level,sub-region level and image level,pixels and their neighborhoods can describe the microstructure of the image;each sub-region can represent the components of the image;the component objects and the background constitute the overall image.The observation and description of the image can be carried out according to the microscopic ? mesoscopic ? macroscopic order,corresponding to the pixel and neighborhood ? sub-region ? image three levels.The traditional local feature extraction technique is usually based on the sub-region description.According to the existing literature,few local feature techniques are based on the microstructure description except of LBP.LBP is a binary description method,which affects its distinctiveness of the target object and recognition application.But the LBP descriptor's defect is that it only preserves the gray-scale relationship between the pixels during the binarization process,and the result in extensive information loss in the image.Based on the simple binary image 3 × 3 neighborhood,this paper presents a novel Binary Image Microorganism Pattern(BIMP)method based on binary image and extends it into gray image.i.e.,Gray Image Micorsructure Maximum Response Pattern(GIMMRP)method.In this method,the image 3 × 3 neighborhood structure is binary coded to obtain the description of the image microstructure,and then the important execution patterns and the pooling operation are selected to realize the representation of the whole image.In order to test the validity of the algorithm in recognition,this paper tests the ORL,YALE two human face data sets,MNIST,USPS two handwritten digital public data sets,and non-public vehicle data sets,showing that the method has strong distinctiveness and robust,and can achieve and exceed the performance of many of the latest algorithms.In order to achieve the rotation object detection,the BIMP pattern is sorted according to eight directions and the statistical histogram is obtained from the segmentation process in the pooling process,which can approximate the rotation invariant description of the feature in order to realize the rotation object detection.The detection of various rotation figures fully verified the effectiveness and feasibility of the algorithm.The method of describing local microstructure and image directly from the original gray image not only has the similar advantages of LBP,but also has strong ability of microstructure description,which greatly improves the recognition ability for local binary description operator.Which has a wide range of applicability in the object detection and recognition field.