Alapfogalmak
The proposed two-branch network architecture can effectively detect both anomalous features and compression artifacts, outperforming state-of-the-art methods on a new challenging image manipulation detection benchmark dataset.
Kivonat
The paper introduces a new Challenging Image Manipulation Detection (CIMD) dataset to evaluate the performance of image manipulation detection methods in challenging conditions. The CIMD dataset consists of two subsets:
CIMD-Raw Subset:
- Evaluates the performance of image-editing-based methods in detecting small manipulation regions across copy-move, object-removal, and splicing forgeries on uncompressed images.
- Ensures each type of manipulation contains the same number of samples for fair evaluation.
- Uses high-quality 16-bit TIFF images to eliminate compression artifacts.
CIMD-Compressed Subset:
- Evaluates the effectiveness of compression-based methods in detecting compression inconsistency using double-compressed images with identical quantization factors (QFs).
- Contains splicing manipulation images where the background is double-compressed while the tampered region is single-compressed, using the same QF from 50 to 100.
The paper also proposes a new two-branch network architecture that can detect both anomalous features and compression artifacts. The model uses HRNet as the backbone and incorporates Atrous Spatial Pyramid Pooling (ASPP) and attention mechanisms to precisely localize small tampering regions.
The frequency stream of the model learns compression artifacts by feeding the image through a novel compression artifact learning module that can detect double compression traces even when the QFs are the same. The outputs of the two branches are adaptively aggregated using a soft selection approach.
Extensive experiments on the CIMD dataset show that the proposed method significantly outperforms state-of-the-art image manipulation detection methods in both challenging scenarios.
Statisztikák
The dataset contains 600 uncompressed TIFF images in the CIMD-Raw subset and 200 JPEG images in the CIMD-Compressed subset.
Idézetek
"To address the issues and challenging conditions, we present a new two-branch IMD network incorporating both the RGB and frequency streams, such that both anomaly features and compression artifacts can be detected in a single framework."
"Our network adopts HRNet (Wang et al. 2020) as a feature extractor, with parallel processing at four different scales as in Fig. 2."
"For the frequency stream, we feed the backbone with quantized DCT coefficients, Q-matrix, and novel residual DCT coefficients from multiple recompressions to detect double compression artifacts."