Handwriting Word Retrieval Algorithms And Applications To Historical Documents Using Deep Learning Method

With the rapid growth of information, the efficient retrieval is becoming increasingly important in today's world. How to do the quick word spotting in precious historical resource, and make full use of this historical information is a major research topic in the field of information retrieval. The technology about information retrieval on one hand can help us quickly find the information that they need and reduce the search pressure on users. On the other hand, it also allows us to quickly understand the previous valuable research work. The traditional optical character recognition technology has made a series of fruitful results in the field of information retrieval, but with the retrieved object become more complex and distorted, this traditional optical character recognition technology failed for future quick and accurate retrieval applications.Based on the current information retrieval algorithms, this thesis will use the deep learning as the main tool, and study the handwriting word retrieval algorithms from matching and recognition respects. The main contributions are as follows:1. Implement the multi-feature extraction of the offline handwritten words images, including the histogram of oriented gradients(HOG), local binary pattern(LBP) and spatial pyramid of HOG(SPHOG).2. The matching algorithm based on the metric learning is achieved. Two different structures: Multilayer Perceptron(MLP) and Convolutional Neural Network(CNN) are used to measure the similarity between two words images. For the CNN structure, two-channel CNN, Siamese network and Pesudo-Siamese network are tried and are compared with each other.3. A non-segmentation retrieval system based on the line level is realized by combining the Convolutional Neural Network and Bidirectional Long Short Term Memory Recurrent Neural Network(BLSTM). CNN is used to automatically learn the temporal feature of text line and BLSTM can recognize the English sentence without segmentation. This system completes the retrieval task from the perspective of recognition.4. Two different dataset are used for experiment: George Washington and IAM offline dataset. We realize both the matching and recognition tasks and compare with other popular algorithms. We also visualize the features which are extracted by the CNN...