Research On Optimization Of Neural Network For Image Matching

The purpose of image matching is to establish the correspondence between two or more images,which is the basis of many computer vision tasks such as augmented reality,simultaneous localization and mapping,3 D reconstruction and so on.It is widely used in national defense military,medical diagnosis and other fields.Image matching in engineering is mainly realized by feature matching technology.Feature matching is generally composed of feature points and descriptors.Feature points are location coordinates of features and descriptors are local descriptions of features.In order to meet the requirements of related tasks,feature matching methods are usually required to have fast processing speed and good robustness to scene changes.The traditional feature matching method obtains the feature points and descriptors through the artificial design algorithm,and outputs the matching results according to the difference of the descriptors.However,the traditional method is mainly based on the local gray level and texture of the image to describe the features,which can not meet the requirements of robustness under the conditions of illumination change and target deformation,therefore,feature matching method based on deep learning has become the focus of researchers.Compared wi th the traditional method,the method based on deep learning can get the feature points and descriptors by pre-training neural network,however,there are still some problems,such as the large number of parameters and computation,and the difficulty of real-time computation in mobile terminal.This paper focuses on this problem.In order to reduce the computational load of neural networks,a gradient adaptive binary quantization method is proposed.In the pre-processing stage,the weights of the pre-quantization network are processed by batch regularization,which improves the information expression ability of the post-quantization network and realizes the smooth quantization process.The quantization scale factor is designed to reduce the quantization error,and in the reverse propagation stage,the gradient loss problem is solved by dynamic scaling of Hyperbolic Tangent function,and the efficient training is realized.Binary quantization reduces the computation of the network,improves the inference speed of the network,but brings the loss of precision.In order to compensate the loss of precision caused by binary quantization,a multi-scale method based on image bit is proposed in this paper.The information distribution of different bits in the image is obtained by extracting and filtering the information according to the bits,and analyzing the information visually.In order to reuse the effective information of image,this paper designs a hierarchical extraction structure based on strictly controlling the quantity of parameters and computation,constructs a multi-scale network of bits,and solves the problem of precision loss caused by quantization.In the experiment of image classification,the performance of the two methods is verified by effective contrast,and the simulation and hardware platform experiments are carried out on the task of feature matching.The experimental results show that the proposed method can effectively solve the problems of neural network computation and real-time operation,and realize efficient and robust real-time feature matching.