A contrastive learning-based heterogeneous dual-branch network for source camera identification

Source camera identification has been a significant focus in image forensics over the past decades. However, as camera model and instance related forensic features are weak compared to image contents, identification performance is far from satisfactory for practical applications. This paper introduces a novel contrastive learning strategy, aimed at enhancing the learning of camera fingerprints by leveraging the similarity between the two branches in a heterogeneous dual-branch network. Initially, a heterogeneous dual-branch feature extraction module is designed, employing two distinct strategies: noise residual estimation and progressive direct estimation, to independently extract forensic information. Contrastive learning is then utilized to enhance shared forensic features related to camera models between the two branches while filtering out irrelevant content residuals. During training, in addition to supervised classification loss, both spatial and frequency losses are applied to ensure the features consistency between the two branches, thereby enhancing the similarity of the features learned by both branches in the spatial and frequency domains. Drawing inspiration from the peak correlation energy metric commonly used in traditional methods, a frequency domain correlation loss is proposed. Extensive experimental results on the Dresden and Vision datasets demonstrate that the proposed method outperforms state-of-the-art approaches. Furthermore, it shows improved robustness against common preprocessing attacks such as JPEG recompression and image resizing, making it more suitable for real-world applications.