Research On Identify Method Of Pornographic Images Based On Sparse Representation

With the rapid development and wide application of Internet, identifying and filtering of pornographic images have became the key technology should be solved urgently, for clearing the net environment and protecting mental and physical health of juveniles. In this thesis, on the base of summarizing of existing methods, an approach is proposed to identify pornographic images based on sparse representation in over-complete dictionary. The statistic difference in high level semantic features are extracted with dictionary learning firstly, then the two reconstructing errors between the input image and its sparse representation in corresponding dictionary are used as classification feature, finally the support vector machine is used as classifier to identify the pornographic images. Due to the most training work are finished at the stage of dictionary learning, the classification features of test image could be extracted quickly and adaptively, avoiding the fine features processing to it, so the classification of unknown image could be finished quickly.The main works are finished in this thesis:(1)The background and significance of this research are introduced, and the existing methods to filter pornographic images information are compared and summarized. The development trend of this research topic is also analyzed.(2) The theory and methods of sparse encoding and representation are introduced in detail, and the learning algorithm of over-complete dictionary of sample images are introduced and fulfilled. In order to select the specific atom to pornographic images, the redundancy analysis of dictionaries of both pornographic images and normal images is proposed, which is conducted by using the Self-Organizing Map artificial neural net.(3) The sample datasets of pornographic images and normal images are created, which are big sampled and multi-typed, and it provides the good data source for improving the reliability of dictionary learning and for subsequent research on filtering the pornographic images.(4) An approach based on sparse representation in over-complete dictionary is proposed to identify pornographic images. The dictionaries of pornographic images and normal images are created respectively by training of big sample images, which are universally applicable. A feature extraction method based on the sparse representation reconstructing errors is given. Finally the experiment evaluation of this algorithm, in which the support vector machine is used as classifier, are finished and the result indicate the pornographic images could be identified effectively with this method, and the classification correct rate is above90%.