Detection and localization of image forgeries using improved mask regional convolutional neural network

Area of Science:

• Digital Forensics
• Computer Vision
• Deep Learning

Background:

• Traditional forgery detection methods rely on handcrafted features, limiting their descriptive power and efficiency.
• Deep neural networks offer superior feature extraction capabilities for complex, high-dimensional data like images.
• Existing deep learning approaches for image forgery detection require improvement in distinguishing subtle manipulation artifacts.

Purpose of the Study:

• To develop a more effective deep learning model for detecting and localizing image forgeries.
• To enhance the discriminative feature extraction for tampered regions in digital images.
• To improve the accuracy and efficiency of identifying copy-move and splicing image manipulations.

Main Methods:

• An improved Mask R-CNN architecture incorporating a Sobel filter in the mask branch was proposed.
• The Sobel filter was integrated as an auxiliary task to align image gradients between predicted and ground truth masks.
• The model was trained and evaluated on standard datasets for forgery detection tasks.

Main Results:

• The proposed Mask R-CNN with an integrated Sobel filter demonstrated superior performance in forgery detection and localization.
• The auxiliary Sobel filter task effectively encouraged the network to learn more discriminative features related to image manipulations.
• The model successfully detected both copy-move and splicing types of image forgeries with high accuracy.

Conclusions:

• The improved Mask R-CNN with a Sobel filter offers a powerful and efficient solution for digital image forgery detection.
• Integrating gradient information via the Sobel filter significantly enhances the model's ability to identify tampered regions.
• The proposed method represents a state-of-the-art advancement in the field of image forensics.