Large-scale Tourism Attractions Image Retrieval

With the rapid development of tourism, various spots pictures have flooded on the web, and the data of image are larger and larger. Many times there are some beautiful scenery pictures that we need to browse, but we do not know which scenic spot it belongs to. Therefore, it is a very practical and urgent problem that how to search the information we desire quickly and accurately from such a large scale data. In recent years, the methods of web-scale image retrieval have become more mature, which provides the possibility of accurately finding a target image from the mass attractions images. In this thesis, combining the web-scale image retrieval methods with image processing methods, we achieve the retrieval of web-scale image of tourist attractions. Our main work is as follows:(1) We study the basic theories and methods of image retrieval, mainly including: GIST descriptor based on image global feature and bag of words model (bag-of-feature, referred to BOF) based on image local features. GIST is that using multiple sets of Gabor filters convolution and image, and the meshing image, and then we get the global feature by cascading the convolution of different grids. However, the feature extraction method usually relies on the division of the grid and has good image retrieval performance only in the case of higher overall similarity different images. BOF model is a reference in the text retrieval method, generally it is the method that extracting image local features by SIFT (scale invariant feature transform) algorithm and using k-means to cluster these features to obtain a low-dimensional visual dictionary, and then we express the image feature by the histogram vector based on visual dictionary. However, when more types of images, the dimension of visual dictionary will be great, and it is not easy to build BOF model for us.(2) In order to improve accuracy shortcomings of two methods in large scale image retrieval, in the thesis we use the framework for feature extraction that 8-layer convolutional neural network Alex made, and take a final full connection layer as an image feature and reduce its dimensionality by principal component analysis, and then we construct a low-dimensional feature index structure by the approximate nearest neighbor algorithm based on local sensitive hash (locality sensitive hashing, referred to LSH). Using the strengths of feature extraction based on convolutional neural network and the hash index structure efficiency in image retrieval, we solve the shortcomings of traditional methods in image retrieval accuracy and other aspects.(3) In the thesis, we use the algorithm to make the image retrieval experiment on 1740 tourist attractions in Beijing, the results show that, the method for most attractions has a clear advantage on the accuracy of retrieval, compared with the above two methods, but when the number of similar images of a scenic spot is small in the attractions image library, the accuracy of query results is not much difference from the first two methods.