Dual-Branch Texture Enhancement Framework for Steganographic Embedding Cost Learning

Cost-based image steganography can significantly enhance its performance through a Reinforcement Learning (RL) framework. However, existing methods still face limitations in the generation of reward signals and the capture of image texture details. To address these challenges, this paper proposes a Dual-Branch Texture Enhancement Reinforcement Learning framework (DBT-RL) for symmetric embedding cost learning. This framework incorporates a Texture Information Enhancement Module (TIEM), enabling the policy network to more effectively focus on complex textured regions. Additionally, DBT-RL integrates multiple steganalysis to construct an ensemble environment network and introduces a novel adaptive update strategy. This strategy dynamically selects the best-performing steganalyzer to provide rewards to the policy network while self-updating weaker steganalyzers, ensuring that the policy network receives precise and dynamically balanced feedback. A large number of experimental results show that DBT-RL achieves high performance in the security of symmetric-cost embedding steganography.