Deep Learning In The Application Of Image Recognition And Algorithm Research

In recent years,deep learning has become a hot research topic in the field of image recognition.In the traditional methods of image recognition,the key of preprocessing is to extract the input data by artificial experience,which waste time and energy in adjusting the parameters in this way.Deep learning simulates the human visual system and the cognitive processes of the human brain,extracts data features by the structure and obtains the deep hidden information of the image.The entire process can get good results without human intervention.The thesis mainly studies the application of deep learning in the field of image recognition.Research the problems of the high accuracy and the structure of the deep learning model.A scheme of the structure of the additive momentum adaptive learning rate method and the construction of the ideal model are proposed in turn.The main works are as follow:1.In the deep learning model training phase,there will be an objective function in local minimum value,a slow convergence speed and the model is easy to appear over fitting and so on.In this thesis,the problems are solved by using the modified additive momentum method and the adaptive learning rate method.Discussing the influence of relevant parameters on the learning process.Experiments using MNIST data set on the stacked denoising autoencoder.The results show that the improved algorithm can improve the accurate rate.At the same time,the training time is shortened,and the feasibility and effectiveness of the improved algorithm are verified in the field of image recognition.2.Deep learning network model requires the designer’s experience and several experiments,which confirms the structure.This approach lacks a reasonable theoretical basis.In this thesis,we propose four methods to determine the reasonable structure of deep learning:the method of parameter analysis,the adaptive golden section principle,The network structure based on the network node sensitivity,and the structure of the Pyramid model.Experimental analysis shows that the four schemes can achieve very good results in recognition accuracy and training time.The network structure based on the network node sensitivity is more excellent.The optimal model structure will provide the theoretical basis for the selection of reasonable structure in the application of the deep learning network in the field of image recognition.