Abstract
Recently, the accessibility and user-friendly nature of image editing tools have increased, allowing even inexperienced users to create and share forged images. Therefore, developing forensic methods to detect forged images is crucial. JPEG image tampering often involves recompression with a different quantization table, known as double JPEG compression. This paper proposes a multi-branch convolutional neural network and compares it with single-branch models to demonstrate its effectiveness in detecting double JPEG compression. The network consists of inter-branches, capturing statistical correlations across all Discrete Cosine Transform (DCT) frequency bands, and intra-branches, focusing on within-band correlations. By increasing feature extraction through additional intra-branches, the system enhances detection performance, particularly in complex datasets with diverse quantization tables. Features are concatenated with the image quantization table to improve robustness across varying quantization table combinations. Evaluated on the Park dataset, which includes over a million JPEG patches and 1,120 randomly assigned quantization tables, the proposed multi-branch model outperforms single-branch architectures (VGG16, DenseNet121, ResNet50) and surpasses state-of-the-art methods with a 94.15% accuracy. Furthermore, it demonstrates superior performance in localizing manipulated regions in real-world images.