The detection of double JPEG compression with the same quantization matrix is a challenging problem in image forensics. In this paper, a CNN framework is proposed to solve this problem. This framework contains a prepr...
详细信息
ISBN:
(纸本)9789881476852
The detection of double JPEG compression with the same quantization matrix is a challenging problem in image forensics. In this paper, a CNN framework is proposed to solve this problem. This framework contains a preprocessing layer and a well-designed CNN. In the preprocessing layer, the rounding and truncation error images are extracted from continuous recompressed input samples and then fed into the following CNN. In the design of the CNN architecture, several advanced techniques are carefully considered to prevent overfitting, such as 1x1 convolutional kernel and global average pooling layer. The performance of proposed framework is evaluated on the public available image dataset (BOSSbase) with various quality factors (QF). Experimental results have shown the proposed CNN framework performs better than the state-of-the-art method based on hand-crafted features.
Detecting double Joint Photographic Experts Group (JPEG) compressionfor color images is vital in the field of image forensics. In previousresearches, there have been various approaches to detecting double JPEGcompress...
详细信息
Detecting double Joint Photographic Experts Group (JPEG) compressionfor color images is vital in the field of image forensics. In previousresearches, there have been various approaches to detecting double JPEGcompression with different quantization matrices. However, the detectionof double JPEG color images with the same quantization matrix is stilla challenging task. An effective detection approach to extract features isproposed in this paper by combining traditional analysis with ConvolutionalNeural Networks (CNN). On the one hand, the number of nonzero pixels andthe sum of pixel values of color space conversion error are provided with 12-dimensional features through experiments. On the other hand, the roundingerror, the truncation error and the quantization coefficient matrix are used togenerate a total of 128-dimensional features via a specially designed CNN. Insuch aCNN, convolutional layers with fixed kernel of 1×1 and Dropout layersare adopted to prevent overfitting of the model, and an average pooling layeris used to extract local characteristics. In this approach, the Support VectorMachine (SVM) classifier is applied to distinguishwhether a given color imageis primarily or secondarily compressed. The approach is also suitable for thecase when customized needs are considered. The experimental results showthat the proposed approach is more effective than some existing ones whenthe compression quality factors are low.
暂无评论