Research On Character Image Recognition And Image Retrieval Based On Deep Learning

Feature extraction and representation of images is a key step in the image recognition and image retrieval.Generally,the excellent feature extraction and representation methods not only bring convenience to subsequent recognition and classification algorithms,but also improve the performance of the entire visual system.However,research on image recognition and image retrieval algorithms based on traditional methods,their classification principle is based on the underlying pixel visual features of the image.This classification method is essentially different from “person” in recognizing and classifying images by understanding image content.Therefore,image recognition and image retrieval systems based on such methods have disadvantages such as: recognition accuracy needs to be improved,features are difficult to migrate and manpower consumption.Although image feature extraction and representation methods based on currently available deep learning shown better results compared with traditional methods,with the increasing demand for improved accuracy and reduced time consumption in the field of image recognition and retrieval in the industry,the methods based on the deep learning image recognition and retrieval will need further research later.Considering above problems and situations,the main work and innovations of this dissertation are as follow:(1)Through the mathematical analysis of the deep learning model,this dissertation sum up the characteristics and advantages of the deep learning methods in the theoretical model.Combined with the research background,we summarized some optimization experience in data preprocessing design,deep neural network framework construction and nuclear model design.(2)We propose an algorithm--Research on a handwritten character recognition algorithm based on an extended nonlinear kernel residual network,which is a deep learning method,and include three main innovation aspects: 1)A new deep learning kernel structure is proposed;2)Design of a semi-supervised apriori algorithm for intra-class clustering;3)Addition of a dropout layer.We conducted experiments on different character image recognition databases MNIST and SVHN,which proved the effectiveness of our method.(3)We propose an algorithm--an image retrieval algorithm based on extended nonlinear kernel residual network and hash method.The method uses the extended nonlinear kernel structure,which proposed in above,as the "high-level semantic extractor" in the image retrieval system,and combines the quick search advantage of the hash algorithm,to complete the fast and accurate retrieval of images.The algorithm was tested on CIFAR-10,which proved the effectiveness of the method.