Multi-label classification for image tamper detection based on Swin-T segmentation network in the spatial domain.

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Current deep learning methods for image forgery detection lack accuracy and localization capabilities.
• Existing models typically support only one type of image tampering, limiting their practical application.

Purpose of the Study:

• To develop a novel deep learning model for accurate and localized detection of image forgeries.
• To enable the detection of multiple image tampering types within a single image.
• To improve upon the performance of existing image forgery detection algorithms.

Main Methods:

• A spatial perception module integrating Spatial Rich Model (SRM) and constrained convolution for focused tampering trace detection.
• A hierarchical feature learning architecture combining Swin Transformer and UperNet for multi-scale tampering pattern recognition.
• A comprehensive optimization strategy involving auxiliary supervision, self-supervised learning, and hard example mining.

Main Results:

• The proposed model demonstrated enhanced performance on the MixTamper and DocTamper datasets.
• Achieved a 13% improvement in the Intersection over Union (IoU) index compared to leading algorithms.
• Successfully detected multiple tampering types from a single image with high accuracy.

Conclusions:

• The developed model significantly improves the accuracy and localization of image forgery detection.
• The novel approach effectively handles multiple tampering types in a single image.
• This work advances the state-of-the-art in deep learning-based image forensic analysis.