Hybrid LSTM and Encoder-Decoder Architecture for Detection of Image Forgeries

Area of Science:

• Computer Vision
• Digital Image Forensics
• Machine Learning

Background:

• Advanced image editing tools enable sophisticated manipulations (copy clone, object splicing, removal), making visual identification difficult.
• Detecting subtle manipulated regions that lack apparent visual artifacts is a significant challenge in digital forensics.

Purpose of the Study:

• To propose a high-confidence manipulation localization architecture for precise segmentation of altered image regions.
• To develop a method capable of identifying various image manipulation artifacts, including those from resampling and splicing.

Main Methods:

• Utilized an encoder-decoder network incorporating resampling features and long short-term memory (LSTM) cells.
• Employed resampling features to capture artifacts from JPEG compression, upsampling, downsampling, rotation, and shearing.
• Incorporated larger receptive fields and frequency-domain analysis for discriminative feature extraction between manipulated and authentic regions.

Main Results:

• The proposed architecture successfully segmented manipulated regions from non-manipulated ones with high precision at the pixel level.
• Rigorous experimentation on three diverse datasets validated the method's effectiveness in localizing image manipulations.
• A large image splicing dataset was introduced to facilitate effective training and improve localization accuracy.

Conclusions:

• The developed high-confidence manipulation localization architecture offers a robust solution for detecting sophisticated image forgeries.
• The method demonstrates significant potential for enhancing the reliability of digital image forensics and content authentication.