Research And Implementation Of Business Model Recommendation System Based On Face Attribute Extraction

In recent years,with the rapid development of the Internet,the total amount of global data has exploded.In order to solve the problem of "information overload",recommendation system has become a key concern in academia and industry,and has been widely used in practice.However,the commercial model agency industry is lagging behind in this wave,and there is an urgent need for an intelligent and efficient business model recommendation system to replace the current outdated traditional model recommendation process.However,building a recommendation system for business model agencies faces a number of difficulties.Firstly,model recommendation system lacks a large amount of multi-user historical interaction data to support it.Secondly,the traditional model recommendation process is highly subjective.So when constructing the system,it need to solve that what kind of objective representation should be used to describe the "temperament" of models,and how to quantitatively measure the similarity of "temperament" between models.Finally,how should the system efficiently obtain the recommendation results based on the similarity of "temperament" ?In the face of the above problems,this thesis researches and implements a business model recommendation system based on face attribute extraction by taking the facial image of models as a breakthrough.The specific work includes the following aspects:(1)Based on the structure of multi-task learning,a face attribute extraction algorithm based on prior attribute grouping is proposed.The algorithm focusses on the construction of multi-task learning network and the design and balance of multi task loss function.Firstly,a priori attribute grouping strategy is proposed as the basis for the construction of sub-task branches in multi-task convolutional neural network.Secondly,the design and balance method of loss functions in multi-task learning network are discussed,and it is verified that the uncertainty weighting is a more appropriate balance method in face attribute extraction.Finally,the experimental results show that the algorithm in this chapter obtains 91.19% accuracy on Celeb A dataset,which is a competitive results and verifies the potential of multi-task learning network structure based on a priori attribute grouping in face attribute extraction.(2)According to the characteristics of face attribute extraction,an improved face attribute extraction algorithm fusing mid-level convolutional features is proposed.Due to the important value of mid-level representation information in face attribute extraction,this algorithm builds the Midlevel Convolutional Feature Fusion Module(MCFFM)and the Task-unique Attention Module(TAM)on the basis of a prior attribute grouping based face attribute extraction algorithm.The former is responsible for efficiently extracting and fusing mid-level convolutional features from the backbone network to compensate for the vanishment of mid-level representation information in the network and enhance sharing features extracted from the backbone network.The latter is responsible for empowering each subtask branch so that it can selectively learn the useful information from shared features,relying on the Attention mechanism,so as to improve the ability of subtasks to learn taskspecific features,reduce redundant information,and enhance the effect of face attribute extraction.The experimental results show that the proposed algorithm based on the above two modules has a strong improvement in the accuracy of classification on Celeb A dataset,with a maximum accuracy of 92.35%,and a lower network parameter amount of 16.8M,which is better than several existing algorithms.At the same time,the proposed algorithm is efficient and easy to implement,which lays a solid foundation for the construction of business model recommendation system.(3)Using the facial image as a starting point,a business model recommendation algorithm based on face attribute extraction is proposed.Firstly,based on the improved face attribute extraction algorithm fusing mid-level convolutional features,the proposed algorithm constructs a face attribute extraction module,which objectively characterizes each model’s temperament by extracting face attribute feature groups,so that the recommendation algorithm can quantitatively analyze the similarity of temperament between feature groups.Secondly,this algorithm builds a recommendation algorithm module,called Face Attribute-Model Matrix based Collaborative Filtering(FAMM-based CF),which can build a co-occurrence matrix with the help of face attribute features,and filter the model database efficiently through the similarity relationship presented by the co-occurrence matrix.Finally,the best recommended results are obtained.At the same time,based on the proposed algorithm,this thesis builds a specific business model recommendation system,which mainly achieves two functions: finding the ideal model for the business,and finding the matching brand for the model.The experimental results show that the business model recommendation system built with the proposed algorithm is efficient,and the first recommendation result of the two functions in the system obtains 4.49/5 and 4.34/5 average satisfaction scores in the user’s subjective evaluation,which indicates that the system has been approved in the application and can effectively replace the traditional business model recommendation process.