【摘要】 目前基于深度学习的图像隐藏方法通常采用编码器-解码器结构，存在编解码网络不耦合以及特征冗余的问题，会导致隐写图像和解密图像的质量下降。为了解决该问题，提出了一种基于可逆神经网络的单图隐藏框架，通过对特征的精细化处理提高模型性能。首先，该方法将图像隐藏和图像提取建模为可逆神经网络的前向传播和反向传播过程，改善了网络的耦合性。其次，使用特征精炼模块对冗余特征进行精细化处理，以提升模型性能。该模块先通过特征筛选模块减少冗余特征，再经过空间-通道自注意力模块精细化图像特征。将秘密图像隐藏至载体图像的高频区域，以增强隐写图像的隐蔽性。实验结果表明，该方法能够有效地平衡图像隐藏中的高容量、不可见性和安全性。更多还原

【Abstract】 This paper proposes a single-image hiding framework based on invertible neural networks, which improves the model performance through the refinement of features. Firstly, the method models image hiding and image extraction as forward propagation and back propagation processes of a invertible neural network, which improves the coupling of the network. Secondly, the redundant features are refined using a feature refinement module to improve the model performance.This module firstly reduces the redundant features through the feature screening module and then refines the image features through the space-channel self-attention module.The secret image is hidden to the high-frequency region of the cover image to enhance the steganographic of stego image.更多还原