【摘要】 在图像分类领域,通过预测图片标签信息以加强图片信息矩阵的稠密度,是解决因信息过载导致图像检索效率低的一种比较流行的思路.本文在该思路的基础上,提出了利用矩阵补全的BPR优化方法来提高图像分类效率的算法.本文先通过矩阵补全将图像学习得到的低维向量空间恢复成高维的图像ID-标签矩阵,再通过贝叶斯个性化排序(Bayesian Personalized Ranking,BPR)算法改进基于矩阵分解的矩阵补全算法,优化标签排序,进而预测图片标签,完成图像分类.本文首先对比了三种矩阵补全模型,分析得到矩阵分解的矩阵补全模型性能最优;其次,在Librec工具库和Google的Open Image数据集上,将改进的奇异值分解算法(FunkSVD)与贝叶斯个性化排序(Bayesian Personalized Ranking,BPR)优化的矩阵分解算法进行对比实验.结果表明,无论是在准确率、召回率、AUC还是计算效率上,BPR优化后的矩阵补全结果值都优于FunkSVD.更多还原

【Abstract】 In the field of image classification,in order to solve the problem of the low efficiency of image retrieval caused by information overload,it is a popular idea to improve the density of picture information matrix by predicting the image label information. Based on this idea,an algorithm to improve the efficiency of image classification by using the BPR optimization method of matrix complement is proposed in this paper. Firstly,the low-dimensional vector space obtained by image learning is restored to the high-dimensional image id-label matrix through matrix complementation,and then improve the matrix completion algorithm based on matrix factorization by Bayesian Personalized Ranking( BPR) algorithm,and optimize the result. Then the image feature label is predicted,classification is completed. this paper first compares three matrix complement models,and demonstrated that matrix factorization model has the best performance. Secondly,the improved Singular Value Decomposition algorithm( Funk SVD) and Matrix factorization algorithm optimized by BPR is compared on the Librec Tool platform and the Open Image dataset of Google. The results showthat the matrix complement results of BPR optimization are better than Funk SVD in terms of accuracy,recall rate,AUC and computational efficiency.更多还原