Application Research Of Convolution Feature Selection In Image Retrieval

Image retrieval is an important branch of computer vision.The general processes of image retrieval are as follows.First,it extract features of the training set image.Second,it extract features form the image which is ready for retrieving.Third,it calculate the feature similarity between the features of first step and feature of the second step and rank the features similarity from low to high.At last,according to the ranked feature similarity,it show the relevant images.Image feature extraction method widely adopts scale-invariant feature transform extraction(SIFT)method.But the main problem of SIFT method is most of the extracted features are underlying features,which have a greater correlation with the underlying texture and contour structure of the image and have a low level semantic relationship with people's understanding of images.With the widely use of deep learning in computer vision,researchers combine deep learning with image retrieval.They use pre-trained convolutional neural network(CNN)to extract features.To some extent,the appearance of deep learning solve the problem that SIFT features cannot express high-level semantics.However,mostly researchers directly use features extracted from CNN.Considering that the CNN model generally uses the classification task as the guidance to train the network weights,so the weights of the trained CNN model may be more suitable for classification tasks than image retrieval.Based on that,we propose Selective Convolution Feature Fusion(SCFF)method and Space Channel Enhancement SCFF(SCE-SCFF)method to solve the problem of feature selection.The main work of the thesis is as follows:(1)Proposed SCFF algorithm.SCFF algorithm is compared with the other two feature fusion methods,one is the max pooling feature fusion(MPFF)method,and the other is the sum pooling feature fusion(SPFF)method.The SCFF method obtains a mask which is two-dimensional plane vector by calculation,and then uses the mask to select features related to image retrieval,and the same time filter out irrelevant features.Finally,we generate a feature vector form the selected CNN feature using SPFF method.(2)Proposed SCE-SCFF algorithm.SCE-SCFF method is further modification and improvement based on SCFF method.Consider that the weight of the selected feature channel may be different,so we refer to the idea of SENet(Squeeze and Excitation Networks)which model CNN channel by weighting.We also construct CNN channel weighting on the selected CNN features.After getting the channel weighting,we do reweight operation which is the same as in SENet.At last,we also generate a feature vector using SPFF method(3)We test SCFF method and SCE-SCFF method on image retrieval dataset which are Paris6 K,Oxford5K and Holiday.We compare our method with SIFT method and other CNN based method in different feature vector dimension.When the image feature dimension is 256,the best result are achieved on the Oxfotd5 K and Paris6 K dataset,and the mean average precision(mAP)are 71% and 79.3%.When the image feature dimension is 512,the best result is achieved on the Oxfotd5 K dataset,and the mAP is 73.0%.