| With the development of Internet technology,the information exponentially exploded,and text search has been unable to satisfy people’s search Demand.Compared with the information provided by the text,the picture has its natural intuitive advantage.Image recognition and retrieval technology has been widely used in various fields,and Shopping with Image retrieval has become popular in e-commerce platform.Users can take photos anytime and anywhere,and search for products quickly in the e-commerce platform.The convolution neural network(CNN)is a representative visual model in deep learning with its local perception and multi-layer network.The special structure plays a great advantage in the processing of image recognition,which is widely used in modern pattern recognition in the field.This paper summarized the theoretical achievements and application status of deep learning and transfer learning.Then we built a new image recognition and image retrieval model based on the deep learning framework of Tensor Flow.And we have practiced it in the e-commerce platform.Firstly,we have studied the research status and the difficulty of image retrieval.Secondly,we got the latest image data from the Taobao platform.After cleaning the data,the images have been rotated,cut,recolored stochastically by data augmentation in order to rich the sample and reduce errors.Then we fine-tuned the neural network model to identify the image combined with the transfer learning method.Through deep learning,the image categories were identified,including category information,dressing style,brand information,etc.Finally we calculated image similarity by using image features from CNN.The accuracy of primary category’s classification was 99.1%,then the second-class and image style were 87.7% and 91.8%.The MAP of image retrieval got 0.64.We can see that the accuracy of our model based on transfer learning is slightly higher than the traditional similarity algorithm from the experimental results. |
PDF Full Text Request