Abstract: The Scale Invariant Feature Transform（SIFT） has been widely used in video concept detection.A lot of researches about SIFT have been done，such as PCA-SIFT，SURF and MESR.But there are few attentions about the influence of different down-sampling ratios to SIFT extraction.Based on the analysis of the influence of different down-sampling ratios to SIFT extraction，and a Multiple-Level SIFT（ML-SIFT） method for senmantic concept detection is proposed.Experiments on Caltech256 and SceneClass13 show that MAPs of ML-SIFT outperform MAPs of SIFT on Caltech256 and SceneClass13 by 15.7% and 5.1% respectively.In addition，when training the models using different ratios of positive and negative samples，the performances of ML-SIFT are stable.At the same time，the comparison of SIFT，SURF and ML-SIFT is given in the paper.From the experiments，the performances of SIFT and SURF are similar，but when comparing to ML-SIFT，their performances are worse than ML-SIFT.From above analysis，the ML-SIFT algorithm is effective.

摘要： SIFT局部特征因良好的性能在图像和视频的语义概念检测中得到广泛应用。已经有很多学者对SIFT做了深入研究，并提出了PCA-SIFT，SURF，MSER等，但是在SIFT算法中，阶与阶之间采样率的变化对SIFT特征的影响关注很少。考察了SIFT算法中，阶与阶之间采用不同降采样率对SIFT特征差异性的基础上，提出了一种多层次的SIFT（ML-SIFT）算法。Caltech256和SceneClass13数据集上的实验表明，ML-SIFT相比于原始SIFT，其MAP的提高能够分别达到15.7%和5.1％。另外在Caltech256上，当采用不同比例的正负样本训练时，ML-SIFT算法具有较好的稳定性。同时，还将ML-SIFT算法、SIFT、SURF算法做了性能比较，SURF和SIFT算法的性能较接近，但是SIFT和SURF相对于ML-SIFT算法，则其性能较差。实验表明，ML-SIFT是有效的。