Abstract: In order to solve the problem that the traditional pornography filtering algorithms are hardly to be used for the complex Internet environment， a novel filtering algorithm was presented based on multiinstance learning by building classification space. Firstly， Hessian matrix was used in YCgCr space to detect image feature points which are instances of the image， and then LBP operator was expanded to YCgCr space. Secondly， YCgCrLBP operator was constructed to describe the image instances. Finally， classification space model based on frequency statistical theory was proposed， and cosine similarity was used to complete image recognition. Different data sets were used to make comparison. The results showed that using the proposed method， the accuracy is increased compared with the large skin contented images filtering by the conventional skin proportional method， and the description of the proposed method is improved compared with the general multiinstance learning method. What’s more， better experimental results were obtained， which indicated the practical value.

Key words: image filtering, multiinstance learning, local binary patterns, Hessian matrix, YCgCr space