Research On Image Retrieval Technology Based On SURF And Convolutional Neural Network

In recent years,with the rapid development of the technology,the number of pictures on the Internet has also increased.The current popular image retrieval methods mainly include image retrieval methods based on artificial features and image retrieval methods based on deep learning.This article researches and innovates these two types of image retrieval methods respectively,proposes an image retrieval method based on SURF(Speeded Up Robust Features)feature points,and an image retrieval method based on Siamese network.The main work and innovation of this article are as follows:(1)The characteristics of the anti-rotation and anti-scaling transform robustness of the existing hash coding methods based on local feature descriptors,a novel image-aware hash coding method based on SURF and clustering algorithm is proposed.After using a perceptual hashing method to retrieve the set of images to be matched,then use SURF feature point matching to select the result.Aiming at the high time complexity of the existing SURF feature point matching algorithm,a novel SURF feature matching algorithm is proposed.Compared with the original matching algorithm,it greatly improves the real-time and robustness of matching.Experiments show that the image retrieval method has good robustness and real-time performance.(2)In the case that the existing manual partial descriptor does not guarantee superior performance under all circumstances,in this paper,we use the Siamese Convolutional Neural Network to learn a similarity function and compare the similarity between images.For the general Siamese network rotation invariance and translation invariability,a pyramidal pooling layer and scale-invariant layer are added to the convolutional layer and the fully connected layer.In addition,added a regular item to the objective function.Experimental results show that the improved network model is superior to the existing Siamese network model in terms of robustness.