Research And Application Of Image Retrieval Method Based On Transfer Learning

In recent years,with the rapid development of information technology,the amount of image data continues to grow explosively,and the relationship between data has become complex.How to complete image retrieval in data set has become an urgent research direction.Image retrieval is to match the same or similar image data from a data set by comparing the characteristics and information of the image itself.In the process of image research of Huizhou building components,it is found that the differences of modeling features of Huizhou building components are relatively small in general,and the differences of detail features are complex,and the sample collection work is not enough,so it is unable to obtain sufficient experimental data,and the problem of insufficient data in the target domain can be solved by the combination of transfer learning and deep learning.In view of the shortcomings of traditional image algorithm and deep learning algorithm in image retrieval application,as well as the researchability of Huizhou building component data set,this thesis carries out the following research and practice:In this thesis,firstly,image retrieval is described in detail,and the related knowledge of image features,as well as the key technologies in image retrieval are introduced,and the evaluation criteria of retrieval methods are discussed.At the same time,the transfer learning method and related deep learning,convolutional neural network are also deeply studied.On this basis,this thesis improves the convolutional neural network through the model-based transfer learning method,and combines with the traditional image matching algorithm to propose three different image retrieval methods.Firstly,an image retrieval method based on feature point matching is proposed.In this thesis,an improved feature point matching algorithm G-AKAZE,is proposed by combining scale space construction based on nonlinear method with grid statistics to remove mismatching.The algorithm is applied to image retrieval,and the experimental results show that this method can effectively improve the time-consuming and accuracy of the algorithm in the experiment of image matching.In the experiment of image retrieval,the accuracy is also significantly improved compared with other algorithms.Secondly,aiming at the problems of insufficient data collection,lack of computing power and time in the application,an image retrieval method based on transfer learning and hash algorithm is proposed.The convolutional neural network model is trained on largescale standard data sets to obtain the pre training model,Then the weight parameters in the pre training results are transferred to the new model by using transfer learning method,and the weight parameters of the model are adjusted by using the corresponding data set of the application scenario through training.The hidden layer is added after the fully connected layer,and the output information of the hidden layer is hashed.Then the similarity is calculated according to the high-dimensional feature vector output from the fully connected layer,and the most similar search results are output according to the similarity ranking.The experimental results show that the retrieval accuracy of this method is better than other methods.The accuracy rate of the most similar 10 samples is 89.1% on cifar-10 data set,which is 0.8% higher than that without hidden layer and hash calculation.The accuracy rate of the most similar one sample is 93.2% and the most similar 10 samples is 84.1% on the data set of Huizhou building components after data enhancement.Based on the former two image retrieval methods,an image retrieval method based on multi-class feature decision fusion is proposed.From two convolutional neural network pretraining models,the weight parameters are transferred to the new model,and the deep features output from the full connection layer of the two new models are extracted for feature fusion.According to the matching algorithm of LBP local binary features,Gabor wavelet features,HSV color features and G-AKAZE feature points,the deep-level feature retrieval results and the low-level feature retrieval results are fused for decision-making,and different weights are assigned through experiments to obtain the most suitable retrieval model for cifar-10 data set,flower-17 data set and Huizhou building component data set.The experimental results show that the best retrieval accuracy of the multi-class feature retrieval model is 89.3%,92.5% and 85.4%.Figure [43] table [10] reference [65]...