The Research Of Image Retrieval Based On Local Feature

With the development of image processing technology and the enhancement of computer system computing capability,machine vision is widely used in object recognition,virtual reality,industrial inspection,robot navigation,aviation and aerospace and so on.Image retrieval is a research branch of machine vision,which is concerned with the retrieval and matching mass digital images.It is the research basis of image stitching,target tracking,motion analysis,object recognition,visual navigation and so on.In recent years,with rapid improvement of the embedded system computing capability and the rapid increase of storage capacity.The embedded systems,such as intelligent terminal and robotics have heavy demand on image retrieval.It has great research value to retrieve the required information from huge amounts of data.In this paper,we take image retrieval on embedded platform as the research content,analyzing and summarizing the existing image retrieval algorithm.Since the computing ability is relatively weak and memory and storage space is limited in embedded system,we propose an improved image retrieval algorithm based on local feature.Firstly,we start from the image feature extraction,analyzing the existing feature extraction algorithm for the robustness of the environmental noise and time cost comprehensively.SIFT(Scale-invariant Feature Transform)algorithm can adapt to the change of image scale,rotation,viewpoint,and can overcome illumination change,but SIFT is time-consuming.Our algorithm optimizes the pyramid generation process.We replaces Gaussian convolution with integral image and mean filter base on the central limit theorem,So getting lower time cost and inheriting the robust of SIFT.The raw image needs to take up a large space,and the usual practice is to compress the image.The image compression will cause the loss of features,so the research of feature extraction for compressed image has more practical value.JPEG(Joint Photographic Experts Group)is the most widely used in the image compression algorithms.In this paper,we optimize the stage of JPEG as for preserving features from compression.The rate distortion optimization is introduced into the image feature preserving algorithm to construct the optimization model of image feature and bit rate.The image is divided into several blocks for optimization based on multi-scale feature distribution.In the image descriptor section,we first compare the weak point and strong point of state of art feature descriptors,and choose SIFT descriptor to do optimization.In the stage of descriptor creation,we reassign the feature block in the log-polar coordinate.The Gauss weight combined with distance weights to improve the robustness of image blur.SIFT descriptor will occupy a large amount of storage space and CPU computing resources when matching massive image data,so we introduce the simulated annealing algorithm to screen out the redundant features in the generated descriptor.The simulated annealing algorithm of machine learning is introduced to screen out the redundant features in the generated descriptors to get high efficiency in image matching stage.As for image retrieval,the binary description has less storage space and less time cost than the traditional SIFT floating-point descriptor.In this paper,we propose a sparse projection learning based binary descriptor,which maps the floating-point descriptor to binary descriptor and uses the Hamming distance to match descriptor.Our binary descriptor can greatly decrease time cost while keeping the distance correlations in the Euclidean space.In the image matching section,the traditional tree search algorithm in high dimensional search will encounter the dimensionality disaster,a sharp decline in performance.In this paper,the search framework bases on Multi-probe LSH(Local Sensitive Hash),using step probe detection.We firstly probe all one-step hash bucket,then detect all two-step hash bucket,which has greatly improved the early search hit probability.So by using the less number of hash table can get higher efficiency than traditional LSH algorithm.Finally,we compare our algorithm with the state of art BOW frame.From ANN point of view,the BOW(Bag Of Word)is a learning hash function.The major differences between the LSH and BOW framework is feature code and feature retrieval method.The experimental results show that through the optimization of each stage,our algorithm is more suitable for embedded platform with limited system resources.