Image Feature Extraction and Retrieval Optimization of Book Pages Based on Convolutional Neural Network

Abstract

Focusing on the feature extraction process of convolutional neural network (CNN), this paper establishes a CNN-based retrieval method of book pages. Then, the pretraining and feature finetuning of the CNN were described separately. The performance of the proposed optimization method was demonstrated through experiments. Considering overall performance and transfer learning capacity, the eight-layer VGG-Fast was selected as the structural framework of our CNN. To train the CNN, it is necessary to gather millions of book page images, and complete the complex task of labeling all these images. Given the excellence of VGG in many transfer learning tasks, this paper chooses to pretrain the CNN with a task-independent dataset. After that, a small book page dataset was adopted to convert the knowledge domain of the CNN from image classification to image page retrieval. In this way, desirable retrieval effects were achieved, without wasting lots of time and energy in collecting and labeling a large book page dataset.