Research On Image Matching Algorithm Based On Binarization And Descriptors

Image matching is a hot research direction in the field of computer vision,and it is also the key step of object detection,image retrieval,3D reconstruction and so on.It finds the key points in the image and generates the feature descriptor of the local region to match the corresponding relationship between the images.In these computer vision tasks,background blur,occlusion,light change and other complex situations often exist.In order to improve the accuracy of complex scenes,it is very necessary to obtain more robust descriptors.Convolutional neural networks have been widely used in multiple tasks of computer vision and have greatly improved performance.In the problem of image matching,in order to obtain more accurate effects,the deep learning method is also used to train to obtain descriptors.However,with the deepening of network structure,although the network performance is becoming more and more powerful,the number of parameters is increasing,which leads to the consumption of a lot of computing resources and memory.Firstly,our method studies the descriptor optimization method based on learning.The existing researches only consider the distance of matching descriptors but not the nearest neighbor relationship in the same category of samples.The margins of the widely used triplet loss function are usually fixed according to experience and do not make full use of the data sample to adjust.Aiming at these deficiencies,this paper proposes neighborhood optimization methods to describe distribution is more even,and using a dynamic soft margin based on the triple loss function,according to the distance of the triples from decision boundary cumulative distribution function to dig more valuable samples,dynamically adapt to the current state of training so as to improve performance.Experimental results on the open data sets Brown,W1 BS and Oxford Affine show that this method can improve the robustness of descriptors and thus improve the image matching accuracy.The image matching method based on deep learning has some shortcomings such as large number of network structure parameters and large amount of computation.Therefore,the image matching method based on binarization is further studied in this paper.Firstly,the binarization method based on information entropy is proposed to improve the information expression ability of the network and reduce the quantization error.Secondly,the multi-scale method based on bit-bit image is proposed to compensate the precision loss of the binarization convolutional neural network.Finally,the binarization method of the neural network is combined with the image matching network.The experimental results on two large data sets,Image Net and CIFAR-100,show that the accuracy of each network structure is improved after the addition of the proposed method.The accuracy of top-1can be improved by 0.3% to 1% on Image Net,and the accuracy of top-1 can be improved by 0.7% to 1.7% on CIFAR-100.Experimental results at Hpatches show that the method achieves about 22% parameter compression and about 48% computational compression under binarization network weight,and the average accuracy loss of descriptor matching is about 0.04 when compared with full precision network.The optimized binary matching network not only reduces the memory footprint and speeds up the reasoning speed,but also effectively controls the accuracy loss.