Research On Specific Image Classification Based On Deep Learning

In recent years,with the rapid development of Internet technology,the explosive growth of information make people find almost all the information they want online.While the Internet has brought great convenience to people,some unscrupulous agencies on the Internet use some horror and bad images to attract people's attention.The spread of these images plagued the majority of Internet users,especially young people.How to effectively filter such images is an important issue.In recent years,deep learning has achieved great success in the field of image recognition.Based on the deep learning technology,this thesis improves the existing deep network so that the model can meet the needs of specific image recognition.First,since the number of pictures belonged to a particular category is far less than the number of pictures that do not belong to the category,there is a problem of sample equalization for the specific image recognition task.Second,since the specific image recognition needs to meet the real-time need,the recognition speed should meet certain requirements.Finally,the picture to be identified may be rotated,so the model needs to be better robust.The main work of this thesis includes the following aspects:(1)Aiming at the problem of sample equalization,a deep network model structure with SVDD(support vector daomain description)to increase the precision of the specific image recognition is proposed.(2)Aiming at the problem of recognition speed,a shallow network based on inception structure to increase the speed of image recognition is proposed.(3)For rotated images,adding data enhancement layer in original network structure to increase the robustness of the model and to improve the recognition accuracy of rotated images is proposed.From the experimental results,the model structure of the SVDD plus depth mode network increases the recognition precision of a particular category of picture,especially when the test sample is not balanced.Inception structure network increases image recognition speed greatly at the expense of a certain accuracy rate.After the data enhancement layer is added,the recognition accuracy of the model for rotated pictures has been significantly improved.