Discriminative feature projection for camera model identification of recompressed images

Camera model identification, which aims to identify the source camera model of the query images, has been well studied in the laboratory environment. Most existing methods regard it as a classification problem and rely on well-designed features that characterize the differences between various camera models. In these methods, however, the accuracy of identification results would suffer from the process of image recompression. The information loss reduces the discriminative ability of the designed identification features and results in a serious accuracy loss. To remedy this shortcoming, we investigate the handicap for accurate source identification when the query image is recompressed and creatively propose a new method Discriminative Feature Projection (DFP) to solve this problem. The proposed method learns a discriminative feature projection that projects the designed identification features into a new feature representation invariant to recompression by minimizing the divergence between recompressed and uncompressed images. We also incorporate two constraints that the discrepancy of different images sources should be large and the latent geometric relations of images neighbors should be preserved into our method to reinforce the discriminative ability. Moreover, we conduct extensive experiments over the public Dresden Image Database. Compared with several state-of-the-art methods on camera model identification, the experiment results verify that DFP can achieve significant accuracy promotion when identifying the recompressed images.