Study On Method For Large-Scale Children Book Page Retrieve Based On Deep Hash Network

The electronic resources of children's books such as animation,game and enhancement of reality enhance the expressiveness and appeal of paper books with the multimedia advantages of combining picture and text、sound and voice.However,there is generally no automatic connection mechanism between paper books and electronic resources,which makes it difficult for users to access the corresponding electronic resources easily,accurately and quickly.This paper considers paper books and electronic resources as a content-based image retrieval problem,studies the method of large-scale early childhood book page retrieval based on deep hashish network,designs and implements a million-scale book page retrieval system,and provides users with a "scanning access" way to accurately locate the corresponding electronic resources.Aiming at the three goals of convenience,precision and instant response,this paper proposes a three-step search framework for"pre-processing"-"feature extraction"-"hash acceleration","and studies are carried out in the following areas:(1)Book page image preprocessing.In order to reduce the negative effect of background and geometric distortion on the retrieval of book page images,a new algorithm of uninteractive image segmentation and geometric distortion correction is proposed.Firstly,using the simple Bayesian method based on the difference of color distribution between book pages and background to classify book pages and background pixels,and the rough position of book pages is determined after clustering.Then,using the rough position of the book initializes the DenseCut image segmentation algorithm,and the precise book page region is obtained.Finally,through perspective transform,the book page area is corrected to rectangle,and geometric distortion is corrected.The experimental results show that the preprocessing method can effectively reduce the negative effect of background and geometric distortion on the retrieval accuracy and has good real-time performance.(2)Image feature extraction based on convolution neural network.Because the book page contains extremely rich visual information,the book page retrieval accuracy is highly dependent on a large number of labeled data sets.Therefore,a new method of feature extraction of book page image based on convolution neural network is proposed.First,pretraining the convolution neural network using the task independent dataset(ImageNet);Then,the knowledge domain of convolution neural network is transferred from image classification to book page retrieval by using a small volume of book page data set.Finally,the output of the intermediate layer of convolution neural network is extracted as the character of book page retrieval.The experimental proves that the image feature extracted by the proposed method has excellent retrieval precision and has great potential for feature compression in the millions of book pages.(3)Book page retrieval acceleration based on deep hash network.The similarity of the high-dimensional floating-point characteristics of two convolution neural network outputs measured by the European distance is large,so book page retrieval on a large scale data set can not meet the immediate response requirement.In this paper,a search acceleration method based on deep hash network is proposed.Firstly,the feature vectors of high dimension are divided into several feature segments using the slice layer.The mapping layer then converts each feature fragment into[[0,1]individual float point eigenvalue;Next,the floating point value of the mapping layer output is converted to 1 bit hash value(0 or 1)using the threshold layer,and the conversion of the high-dimensional float point eigenvalue to low-dimensional hash code is completed;Finally,using the hamming distance to calculate the similarity between two hash codes.The experimental results show that the distance measurement time of traversing a million dataset is only 0.33s after the feature vector is converted to 64 bit hash code.Based on the above research,a large-scale book page retrieval and electronic resource association system is realized.Experiments show that the system has a Top-1 hit rate of 82.55%、a Top-5 hit rate of 88.67%and a search speed of 0.61 seconds per page on a single server(one NVIDIA 1080Ti GPU)in a book page data set of 1.55 million.Using this system,users can accurately locate books and access the corresponding electronic resources in an unconstrained manner.