In recent years, various chaos-based image encryption algorithms have been proposed to meet the growing demand for real-time secure image transmission. However, chaotic system that is the core component of chaos-based cryptosystem usually degrades under finite computing precision, causing many security issues. In this paper, a novel cryptosystem with analog-digital hybrid chaotic model is proposed. Firstly, the analog Chen chaotic system and the digital Logistic map are adopted to depict the capability of the hybrid model, in which analog system is used to perturb digital system. Dynamic analyses demonstrate that the hybrid method has better complexity, larger chaotic parameter range and good ability to counteract dynamical degradation. The chaos-based key streams generated by the perturbed Logistic map are more suitable for image encryption. Secondly, a parameter selection mechanism is introduced to increase security. The state variables of Chen chaotic system and cipher image are involved in parameter selection process to dynamically change the parameter of the perturbed Logistic map. The involvement of cipher image makes the key streams relevant to plain image and can resist known/chosen-plaintext attacks. Performance, security and comparison analyses indicate that this cryptosystem has high security, low time complexity, and ability to resist common attacks.
针对神经网络在进行图像着色时容易出现物体边界不明确、图像着色质量不高的问题,提出结合Pix2Pix生成对抗网络的灰度图像着色方法.首先改进U-Net结构,采用8个下采样层和8个上采样层对图像进行特征提取和颜色预测,提高网络模型对图像深层次特征的提取能力;然后使用L1损失和smooth L 1损失度量生成图像与真实图像之间的差距,对比不同损失函数下的图像着色质量;最后加入梯度惩罚,在生成图像和真实图像分布之间构造新的数据分布,对每个输入数据进行梯度惩罚,改变判别器网络梯度限制方法,提高网络在训练过程中的稳定性.在相同实验环境下,使用Pix2Pix模型和summer2winter数据进行对比分析.实验结果表明,改进后的U-Net和使用smooth L 1损失作为生成器损失可以生成更好的着色图像;而L1损失能更好地保持图像结构信息,使用梯度惩罚可以加速模型的收敛速度,提高模型稳定性和图像质量;该方法能更好地学习图像的深层次特征,减少图像着色模糊现象,在有效地保持图像结构相似性的同时提高图像着色质量. 相似文献
