Distinguishing Between Natural and Computer-Generated Images Using Convolutional Neural Networks

Quan, Weize; Wang, Kai; Yan, Dong‐Ming; Zhang, Xiaopeng

doi:10.1109/tifs.2018.2834147

Cited by 102 publications

(93 citation statements)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…BN, on the other hand, requires coarse level information about image sizes to separately process images of different size for better performance. We show that by leveraging on pre-trained models, the finetuning of the proposed networks not only converges much faster than recent works such as [4,7] but also generalizes well for other tasks like photorealism detection of heterogeneous origin [10].…”

Section: Introductionmentioning

confidence: 73%

“…4. Secondly, the CNN model proposed in [4,7] are better suited for resampling detection task and a few other tasks, while the one proposed in [10] is better suited for photorealism detection. However, by building on top of already existing state-of-the-art architectures helps easy generalization to different problems.…”

Section: Results On Photorealism Detectionmentioning

confidence: 99%

“…For example, by using ResNet-18 [17] as base architecture, IPN achieves better performance for both resampling and photorealism detection tasks as corroborated by Table 3 and Table 5. For Table 5, the exact procedure followed by [10] on the Columbia dataset [21] was used to re-train the listed CNN models, except for one change. The Natural Images (NI) and Computer Generated Images (CG) were not resized to a fixed dimension before patch selection.…”

Section: Results On Photorealism Detectionmentioning

confidence: 99%

“…Since the accuracy is contingent on the sharpness of peak, it is obsevred the performance of the method quickly degrades for patches less than 512 × 512. [3] 1024 × 1024 512 × 512 MISLnet [7] 256 × 256 256 × 256 Li et al [14] 512 × 512 512 × 512 Verma et al [9] 512 × 384 128 × 128 Kirchner et al [15] 1024 × 1024 512 × 512 Bianchi et al [2] 1024 × 1024 512 × 512 Quan et al [10] <1024 × 1024 233 × 233 kernel size proportional to tensor's height and width to reduce all tensors to the same size. However, in our experiments this resulted in poor performance, see results for the Max-Pooling Network (MPN) in Table 3.…”

Section: Iterative Poolingmentioning

confidence: 99%

“…Inspired by the immense success of works like Faster RCNN [11] and YOLO [16] which decided to build on top of existing CNN models, we also chose to use ResNet-18 [17] as our base architecture. Although proposed much earlier than recent CNN models for image forensic [4,7,10] Fig. 2.…”

Section: Iterative Poolingmentioning

confidence: 99%

See 4 more Smart Citations

Multi-Patch Aggregation Models for Resampling Detection

Lamba

Mitra

2020

ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

Images captured nowadays are of varying dimensions with smartphones and DSLR's allowing users to choose from a list of available image resolutions. It is therefore imperative for forensic algorithms such as resampling detection to scale well for images of varying dimensions. However, in our experiments we observed that many state-of-the-art forensic algorithms are sensitive to image size and their performance quickly degenerates when operated on images of diverse dimensions despite re-training them using multiple image sizes. To handle this issue, we propose two novel deep neural networks -Iterative Pooling Network (IPN), which does not assume any prior information about the original image size, and Branched Network (BN), which uses this prior knowledge to produce better results. IPN adopts a novel iterative pooling strategy that converts tensors of multiple sizes to tensors of a fixed size, as required by deep learning models with fully connected layers. BN alternatively adopts a branched architecture with dedicated pathways for images of different sizes. The effectiveness of the proposed solution is demonstrated on two problems, resampling detection and photorealism detection, which are generally solved as independent problems with different deep learning models. The code is available at https://github.com/MohitLamba94/ Iterative-Pooling.

show abstract

Section: Introductionmentioning

confidence: 73%

Section: Results On Photorealism Detectionmentioning

confidence: 99%

Section: Results On Photorealism Detectionmentioning

confidence: 99%

Section: Iterative Poolingmentioning

confidence: 99%

Section: Iterative Poolingmentioning

confidence: 99%

See 3 more Smart Citations

Multi-Patch Aggregation Models for Resampling Detection

Lamba

Mitra

2020

ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

show abstract

Weakly Supervised Adaptation to Re-sizing for Image Manipulation Detection on Small Patches

Darmet

Wang

Cayre

2020

Digital Forensics and Watermarking

Self Cite

View full text Add to dashboard Cite

Basic image processing operations like median filtering and Gaussian blurring in general do not change the semantic content of an image, although they are commonly used to cover fingerprints of falsification that does alter image content such as copy-move and splicing. Therefore image forensics researchers are interested in detecting these basic operations. Some existing detectors track local inconsistencies in statistics of the image. However these statistics are very sensitive to image development process. Thus pre-processing operations can be damaging for performances of such detectors. In this paper, we focus on a very common pre-processing operation, i.e., re-sizing, and study how it affects performance when trying to detect several image processing operations on small patches, with Gaussian Mixture Model (GMM) as feature extractor and a Dense Neural Network (DNN) as classifier. We first show performance drops. We then introduce an adaptation method which relies on better fit to testing data for the feature extraction and fine-tuning for the neural network classifier. Experimental results show that our method is able to improve results with very few labeled testing samples. We also present comparisons with an improved version of a recent CNN(Convolutional Neural Network)-based method.

show abstract