Improvement of Image Universal Blind Detection Based on Training Set Construction

The detection rates of existing universal blind detection reduced greatly in practical applications due to the generalization problem. According to the principle of orthogonal design, this paper builds three sample sets of embedding rates mismatch, embedding algorithms mismatch and image sources mismatch between the training sample and the testing sample. The three sets are used to test the detection error rates of Rich Model in the case of embedding rates mismatch, embedding algorithm mismatch and image source mismatch. This paper proposes several methods to improve the generalization ability of the universal blind detection, including training the sample by small embedding rates, learning various kinds of embedding algorithms, pre-classifying the testing sample and improving the IQM algorithm. The results show that the practicability of the universal blind detection will be improved.