Image tampering detection based on RDS-YOLOv5 feature enhancement transformation

Area of Science:

• Computer Vision
• Digital Forensics
• Artificial Intelligence

Background:

• Malicious image tampering poses significant threats to social stability and personal safety.
• Existing tampered detection technologies face limitations in generalization and efficiency due to dataset specificity and feature extraction.
• There is a need for robust and accurate methods for detecting tampered images.

Purpose of the Study:

• To propose an improved tampered image detection method using RDS-YOLOv5 with feature enhancement.
• To enhance the detection of tampering traces through a multi-channel feature enhancement fusion algorithm.
• To improve the robustness and performance of tampered image recognition models.

Main Methods:

• A multi-channel feature enhancement fusion algorithm was developed to highlight tampering artifacts.
• An improved deep learning model, RDS-YOLOv5, was designed for tampered image recognition.
• A nonlinear loss metric for aspect ratio was integrated into the SIOU loss function for optimized training.
• RDS-YOLOv5 was trained using a fusion of original and enhanced image features.

Main Results:

• RDS-YOLOv5 demonstrated performance improvements over the original YOLOv5 model, with gains of 6.46% in F1-Score, 5.13% in mAP50, and 3.15% in mAP95.
• The integration of the SRIOU loss function enhanced the model's ability to locate tampered regions by 2.54%.
• Training with a fused dataset further boosted the overall detection performance by approximately 1%.

Conclusions:

• The proposed RDS-YOLOv5 method effectively enhances tampered image detection capabilities.
• The feature enhancement and improved loss function contribute to a more robust and accurate detection system.
• This research offers a promising advancement in digital forensics for identifying malicious image manipulations.