【摘要】 为解决卷积神经网络(convolutional neural network, CNN)中参数不具有泛化性的问题，融合深度学习和智能算法设计了一种基于卷积神经网络和随机漫步理论的可训练端到端图像去雾算法。首先，利用基于卷积神经网络的可训练端到端图像去雾算法(DehazeNet)计算图像大气透射率；然后，使用K-means算法对大气透射率进行聚类分析，使大气透射率在某一范围内的分布更均匀；接着，利用均方误差函数与失真度函数的差作为优化大气透射率的目标函数，用聚类后的大气透射率作目标函数的初值，利用随机漫步算法求解最优大气透射率；最后，恢复出清晰无雾的图像。实验表明，算法的去雾效果优于DehazeNet算法的去雾效果。更多还原

【Abstract】 The parameters involving image dehazing algorithm based on convolutional neural network couldn’t be generalized. In order to solve this problem, this paper provides a trainable end-to-end image dehazing algorithm by convolutional neural network and random walk. This algorithm combines integrating deep learning and intelligent algorithm. Firstly, this algorithm calculates the image atmospheric transmittance using the trainable end-to-end DehazeNet image dehazing algorithm. Secondly, this algorithm uses K-means algorithm to cluster the atmospheric transmittance. The purpose of this step is to make the atmospheric transmittance more uniformly distributed in a certain range. Thirdly, the difference between the mean square error MSE function and the distortion function is chosen as an object function. Random walk algorithm is adopted to compute the optimum atmospheric transmittance. Finally, a clear and fog-free image is recovered by the atmospheric scattering model. The experimental results show that the dehazing algorithm integrating deep learning and intelligent algorithm is superior to the DehazeNet algorithm in the area of image dehazing.更多还原