Recommendation Research Based On Image Semantics

The Internet has been developing for decades,The information on the Internet also has the original text,voice,picture,and video of the original text information.The proportion of multimedia information to the whole information of the Internet is also getting higher and higher,However,text retrieval based methods can only retrieve little text information on the Internet,At the same time,with the development of Internet blowout users,it is becoming more and more difficult to retrieve information that they are interested in.in this paper,the image information is difficult to be retrieved effectively and the mass data is difficult to be obtained by the user.A recommendation based on image semantics is proposed.The research contents of this paper can be divided into two parts.First,how to effectively acquire image features and further get semantic information of images;second,how to effectively recommend corresponding information to interested users after obtaining image semantic information.In the acquisition of image features,the development trend of convolution neural network is summarized: 1.convolution core miniaturization and model depth.2.increase the width of the model by parallel convolution kernel of different sizes to obtain the image characteristics as much as possible and reduce the complexity of the model parameters.By optimizing the model structure on the basis of the Inception-ResNet-v2 model,So as to improve the ability to obtain image features.In the part of image semantic acquisition,the development characteristics of recurrent neural network are summarized: Internal division of division of labor in the model and the interaction of subunit models.Modifying the model structure of its input gate,output gate,forgetting gate,and update gate on the basis of the long and short term memory model,To improve the performance of image semantics.In the image recommendation section,Three problems,such as cold start,new image cannot be recommended by collaborative filtering,and the same image deweighting mechanism should be faced to the user data recommendation algorithm.The improvement measures are put forward: First,to solve the cold start problem,this paper proposes a new hot content mixed recommendation method to solve the cold start problem.At the same time,a set of evaluation algorithm is put forward to balance the new heat content.Second,in view of the problem that the new image without user behavior can not be collaborative filtering recommendation,this paper uses content based recommendation and collaborative filtering recommendation to superpose the problem,At the same time,the environmental factors are introduced to recommend different contents according to the location,time,and weather information of the user,To better improve the effectiveness of the recommended and user experience.Third,aiming at the problem of the same image deweighting mechanism,this paper proposes a dual factor algorithm of image heat and time to preserve the highest score.Finally,in the experimental section,it is divided into two experiments.The first evaluation is to obtain the semantic experiment of the image.Using the Microsoft COCO dataset and the evaluation criteria it provides to verify that the modified model is improved in performance compared to the previous one,Second experimental results are evaluated using 5 performance indicators,Precision rate,Recall rate,F-measure,ROC curve and AUC area,to evaluate the improved algorithm.The performance has been greatly improved by comparing with previous collaborative filtering algorithms used alone.