Image Retrieval Based On Deep Convolution Features

With the development of multimedia and internet technology,more and more images are used in our daily life.It is quite useful to retrieve the interested images from image databases and has great demand in daily life.Extracting features from images is a key technique in image retrieval.The feature extraction methods based on convolutional neural network have been prevailed due to their high accuracy in image retrieval.Among these methods which are based on the convolution layers,the part-based weighting and aggregation(PWA)method has achieved the best results.PWA first selects some discriminative channels from the convolution features as part detectors to generate spatial weights.Then,the convolution features are weighted by these spatial weights.Next,the convolution features are sum aggregated,which are connected to form the final image representation.Since the method uses discriminative channels,the final feature representation of images contains rich semantic information,and high retrieval accuracy can be achieved.It is noted that,PWA method results in visual burstiness where a number of visual elements are repeated,which makes the features respond suddenly after aggregation and leads to the problem that the burst response dominates the similarity measurement of features and decreases the image retrieval accuracy.To deal with this problem,several methods including power normalization,aggregation based on the L2 norm,channel weighting based on the sum aggregation are proposed.These methods reduce the large responses and increase the small responses,which balances the feature distribution,thereby improves the accuracy of image retrieval.The channel selection in PWA plays an important role in the final extracted features.To cope with the problem that some channels near the parameter threshold are not distinguishable such that the selected channels are not optimal,a channel weighted convolution feature is proposed to select the optimal channels.This method can increase the contrast among channels,facilitates the selection of good channels,and hence improves the accuracy of image retrieval.In addition,the relationship between the variance of the feature map and the semantic information is analyzed,and a channel selection method based on the cumulative variance of the feature map is proposed,which can also help obtain discriminative channels and improve the accuracy of the original PWA approach.Finally,an integrated image retrieval method,namely,the BSSPWA method,which combines the aforementioned two methods is presented.The proposed BSSPWA method firstly uses the convolution layers weighted by the sum-channel weights to select the discriminative channel as the part detectors,and generating the spatial weight by the detectors.Then the spatial weight and the channel weight are both used to weight and aggregate the deep convolution features to form the final image representation.Experimental results performed on four public databases,which are benchmarks in image retrieval,demonstrate that the proposed method greatly improves the retrieval accuracy in comparisons with the state-of-the-art methods.