Efficient Source Camera Identification with Diversity-Enhanced Patch Selection and Deep Residual Prediction

Area of Science:

• Digital Image Forensics
• Computer Vision
• Machine Learning

Background:

• Source camera identification is crucial in digital image forensics.
• Conventional methods struggle with small image patches and varying content.
• Deep learning offers potential but faces challenges in robustness and patch-level accuracy.

Purpose of the Study:

• To propose an efficient patch-level source camera identification method using convolutional neural networks.
• To improve robustness and reduce training costs through diverse patch selection.
• To mitigate the impact of image content on identification accuracy.

Main Methods:

• Utilized a convolutional neural network (CNN) for patch-level camera identification.
• Implemented a representative patch selection strategy for enhanced training data diversity.
• Developed a fine-grained multiscale deep residual prediction module to reduce scene content influence.
• Employed a modified VGG network for identification at brand, model, and instance levels.
• Proposed a critical patch-level evaluation protocol for fair comparison.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art algorithms.
• Achieved improved robustness and accuracy in source camera identification, particularly at the patch level.
• Experimental results validate the effectiveness of the multiscale residual prediction module and patch selection strategy.

Conclusions:

• The developed CNN-based method offers an efficient and robust solution for patch-level source camera identification.
• The proposed techniques effectively address limitations of existing methods concerning image content and patch size.
• The study provides a valuable advancement for real-world image forensic applications.