Research And System Implementation Of Vertical Glasses E-commerce Recommendation Method Based On Face Recognition

Myopia is listed as one of the three major diseases in the world.For myopia patients,glasses are a daily necessity.With the development of economy,people not only appreciate the optical function of glasses,but also emphasize their decorative function.Nowadays,there are so many styles of glasses,which interfere with users'shopping choices.On the other hand,the rapid development of Internet e-commerce has led to the combination of the traditional eyewear industry and new business models,resulting in massive amounts of data.This allows researchers to mine useful information through machine learning and other technologies,generate personalized recommendations,and help users make purchases.select.This thesis takes the glasses e-commerce recommendation system as the main research object,discusses a new recommendation model,and combines user face data and shopping behavior data to achieve accurate and real-time product recommendation.First of all,this paper adopts a collaborative filtering model combined with the small batch gradient descent algorithm MBGD,and improves the problem of fixed learning rate in MBGD,and proposes an adapter-MBGD algorithm to improve the accuracy and real-time performance of the recommendation results;then introduces similarity-based A product recommendation method for face matching,and aiming at the feature of near-sighted people wearing glasses,a occluded face recognition algorithm is proposed to reduce the impact of occlusions on face matching.The recommendation method based on similar face matching provides another dimension of product recommendation,which can solve the problem of low coverage of collaborative filtering recommendation model and weak ability to discover long-tail products,break the deadlock of hot sellers and less popular products,and improve recommendation The robustness of the system;finally design and implement the glasses e-commerce recommendation system.The specific work is as follows:(1)In the recommendation system,the SVD matrix decomposition algorithm is one of the commonly used algorithms.Due to the large amount of data in the e-commerce system,SVD often needs to be used in combination with incremental algorithms.According to the actual situation,mini-batch gradient descent(MBGD)algorithm and SVD algorithm are selected to complete the recommendation.However,the learning rate of the original MBGD algorithm is fixed,which makes the selection of the learning rate difficult and the convergence process fluctuates.In this thesis,for the problem of fixed learning rate of MBGD algorithm,an adapter-MBGD algorithm is proposed,which introduces a second-order stochastic Heun numerical method to approximate the optimal solution.Stochastic Heun numerical method is used to construct the approximation of parameters.Through this approximation,the behavior of the function in the vicinity of the parameter points can be inferred,so as to design independent adaptive learning rates for different parameters.Finally,the attenuation of the stable learning rate is controlled by super parameters,so as to realize the adaptive learning rate.The algorithm ensures that the learning rate keeps stable while accelerating,and effectively solves the problem of MBGD single learning rate.And the algorithm is applied to the SVD model to optimize the recommended algorithm model.Experimental results show that the algorithm proposed in this thesis shows the best result with an RMSA value of 0.949,which is the best choice for matrix decomposition.In addition,the model update time proposed in this thesis is 15%higher than SGD algorithm,which is the fastest update time among the existing models.Therefore,the proposed model is more accurate and real-time.(2)In the glasses e-commerce recommendation system,the face images uploaded by users are often accompanied by mixed noise-structural noise caused by occlusion such ans glasses,and sparse noise caused by the quality of picture shooting,etc.The mixed noise brings great interference to face recognition.This thesis proposes a matrix regression model based on weighted Schatten-p norm and tree-shaped L_?norm(WSTL_?MR)to deal with mixed noises in face images,and uses a distribution model to characterize structural noise and sparse noise respectively.On the one hand,the generalized matrix variable slash(G.M.S.)distribution is used to represent structural noise,and a weighted Schatten-p norm is introduced to solve the rank function problem in the process of deriving the model using maximum posterior probability estimation(MAP).This norm is the best existing solution to the rank function.It is more flexible when dealing with practical problems and at the same time better approximates the original low rank hypothesis.On the other hand,the tree structure norm is used to constrain the sparse noise.The tree-shaped L_?norm breaks the assumption of the L₁norm's independence between pixels and enhances the universality of the model.Based on the alternating direction multiplier method(ADMM),the solution process of the proposed model is given.Experimental results show that the face reconstruction error rate of this model is the lowest,which is 2.11%.Even for real distinguish such as glasses and scarves,this model shows good performance.The recognition rate is improved by 3.25%on average compared with commonly used algorithms.After superimposing sparse noise on this basis,the recognition rate is increased by an average of 10.05%.It has good performance for face recognition under mixed noise.(3)On the basis of the above work,this thesis also completed the design and implementation of the glasses e-commerce recommendation system,realized the glasses try-on function,and combined face recognition and user behavior for personalized product recommendation.