Fast Image Retrieval Based-on CNN Feature Via Hashing

Content-based image retrieval has attracted lots of research effort in the past few years and is probably the most rapidly developing application of similarity search in the 21 st century.In contrast to text-based image retrieval using text labels,content-based image retrieval matches similar images based on image content.When retrieving,the images are compared in the feature space after representation.Traditional content-based image retrieval uses basic image information such as color,texture,shape,edge,etc.to construct image feature.The feature dimension is usually high and the characterization is weak,which cause the retrieval performance suffering from the notorious “curse of dimensionality” and semantic gap.Hashing technology can generate compact and low-dimensional hash codes from sparse high-dimensional image features,which is of great significance to be used in a large-scale image retrieval system.Deep learning makes it possible to generate a highly compact hash code by learning from the high-level CNN feature.Learning to hash takes the advantage of deep neural network to learn a set of hash functions from image representations and challenges the traditional hash methods.The learned hash codes are usually far superior to the simple codes generated by hand-crafted hash functions in performance.In this thesis,we proposed a new framework to generate compact and low-dimensional hash codes by learning from images.The framework uses the convolutional neural network to extract the high-level image features.The hidden hash layer is then used to learn the hash functions based on the features.We took the whole network training as a classification task with the hash function learning task embedded.By fine-tuning on the soft-max loss and quantization loss,the network is tailed to learn hash oriented image feature and generate hash codes which are suitable for fast image retrieval.Moreover,the trained model produces lower quantization error,and our framework is end-to-end trainable.Experiments on MNIST and CIFAR-10 datasets demonstrated the efficiency of our method.