| Along with the advance of all kinds of photography and video technology subject and the rapid development of Internet,images are becoming more and more important in people's life.In this case,the subject of image retrieval was born.However,it has been a difficulty to find a image which meets the need of users quickly because of the massive growth in the number of images.In order to explore the methods to improve the efficiency of image retrieval,this thesis is mainly based on the convolutional neural network and feature fusion technology.In the first step,we classify all the images from the image library,and then we put the target image into the network which has been trained well to get one or more categories which the target image belongs to.The second step,we do small scale image retrieval among the categories we have got.In terms of image classification,this thesis improves the existing traditional convolution neural network,and chooses the Relu function to be the activation function.Besides,this thesis adds DropConnect function to reduce the probability of overfitting.Three different structures of convolutional neural network are established.Compare the results of the classification of them so as to determine the optimal model.After that,put the target image into the model as a test set,and get the classification result.And then,relying on the probability of each type,choose one or a few classes the target image may belongs to for next step.In this thesis,the image classification has achieved a high accuracy rate.In terms of image retrieval in a small scale,the key steps are feature extraction and similarity matching.In order to improve the deficiency of single feature description,this thesis fuses the HSV color histogram feature and texture feature.Because of that the dimension of the feature vectors is too big,this thesis uses PCA(Principal Component Analysis)to reduce them.Firstly,reduce both features respectively,and then reduce their fusion feature again.In this way,the efficiency of operation can be improved.In the similarity matching,this thesis determines the Manhattan distance as the optimal distance measurement method by comparing the retrieval results of the three metrics.The experiment has achieved good results.It proves the feasibility of the method in this thesis. |
