Related Experiment Videos
DeepMIH: Deep Invertible Network for Multiple Image Hiding
IEEE Transactions on Pattern Analysis and Machine Intelligence
|January 10, 2022
Summary
This study introduces DeepMIH, a novel framework for multiple image hiding using invertible neural networks. DeepMIH achieves superior invisibility, security, and recovery accuracy for hidden images compared to existing methods.
Area of Science:
- Computer Vision
- Digital Image Processing
- Machine Learning
Background:
- Multiple image hiding is challenging due to potential contour shadows and color distortion.
- Existing methods struggle with high-capacity hiding while maintaining image quality.
Purpose of the Study:
- To propose a novel framework, DeepMIH, for high-capacity and high-fidelity multiple image hiding.
- To improve the invisibility and security of hidden images.
- To achieve perfect recovery of all secret images.
Main Methods:
- Developed an invertible hiding neural network (IHNN) for reversible image concealment and revelation.
- Cascaded IHNNs to accommodate multiple secret images.
- Integrated an importance map (IM) module to guide hiding based on previous results.
- Proposed a low-frequency wavelet loss to optimize hiding in high-frequency sub-bands.
Main Results:
- DeepMIH significantly outperforms state-of-the-art methods in invisibility, security, and recovery accuracy.
- The proposed IHNN framework effectively models reversible image hiding processes.
- The IM module and wavelet loss enhance hiding performance and invisibility.
Conclusions:
- DeepMIH provides a flexible and effective solution for multiple image hiding.
- The framework demonstrates superior performance across various datasets.
- Invertible neural networks offer a promising direction for advanced steganography techniques.