Research On Cleaning Image Data Based On Deep Learning

In recent years,with the fast development of information society,artificial intelligence gradually penetrates to every aspect of our lives.And then a new model with stronger ability of learning and of summarizing image characteristics,based on the Convolutional Neural Network(CNN)theory are designed and developed to solve the problems of traditional image recognition.On the one hand,the new method makes great progress in target detection,face recognition,image recognition and even natural language processing.On the other hand,the industry of data-cleaning related products and services booms along with the explosion of Internet information.However,little research on the image data cleaning has been conducted.Hence,to produce high quality dataset and achieve a better effect of CNN training,it is significant for us to make investigations on internet image data cleaning.After a brief introduction of CNN and an analysis of data cleaning method,the rest of this paper presents the main work from these following aspects:(1)Re-examine the current image data set: utilizing the identical CNN to optimized these randomly trained and tested data set made up from different image data sets,and the final test accuracy of CNN should be referenced.(2)Re-polish low recognition rate image data: the CNN model trained by an image data can also be made of a CNN classifier to identify single image.We can discover and clean the low recognition rate images with the probability of every image recognized as itself.In addition,we can get and clean the useless image categories which take small part of image data according to the amount of each kind of image.(3)Re-assure the cleaning result by comparison: The final data set cannot be differentiated from the original one due to their dissimilarities in categories and numbers.Thus,a comparative experiment is carried out to establish a set of different training datum belonging to the same classes to evaluate the same testing data set.Then we can draw the conclusion from the result.(4)From the same image data set,experiments based on AlexNet and GoogLeNet showed the improvement of CNN test accuracy respectively up by 1.5% and 2.4% after using the cleaning method proposed by this paper.So,the method in internet works out in image data cleaning.(5)An online flower image recognition system based on the model of our cleaning result is built,so that other users can make online flower picture recognition through simple operation of their Android mobile phone.