Based on convolutional neural network (CNN), the problem of robust patch level camera model identification is studied in this paper. Firstly, an effective feature representation is proposed by concatenating a multiscale residual prediction module as well as the original RGB images. Motivated by exploration of multi-scale characteristic, the multiscale residual prediction module automatically learn the residual images to avoid the subsequent CNN being affected by the scene content. Color channel information is integrated for enhanced diversity of CNN inputs. Secondly, a modified richer convolutional feature network is presented for robust camera model identification by fully exploiting the learnt features. Finally, the effectiveness of the proposed method is verified by abundant experimental results at the patch level, which is more difficult than image level experiments.
|Title of host publication||2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference|
|Publication status||Published - Nov 2019|
|Event||2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019 - Lanzhou, China|
Duration: 18 Nov 2019 → 21 Nov 2019
|Conference||2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019|
|Period||18/11/19 → 21/11/19|