摘要： 高强度的高斯噪声往往会模糊或破坏图像的细节和结构，导致边缘信息的丢失。为此，提出基于逐像素强化学习的边缘保持图像复原算法。首先，为每个像素构建一个像素层智能体并设计针对边缘处的侧窗均值滤波器到动作空间中，所有的像素层智能体通过共享优势行动者-评论家算法的参数，模型可以同时输出所有位置的状态转移概率并选择合适的策略进行状态转移，从而复原图像；其次，在特征提取共享网络中结合协调注意力，从而聚焦所有像素位置在特征通道间的全局信息，并保留位置嵌入信息；然后，为了缓解稀疏奖励问题，设计一个基于图拉普拉斯正则的辅助损失，关注图像的局部平滑信息，对局部不平滑区域加以惩罚，从而促进像素层智能体更加有效地学习到正确的策略来保持边缘。实验结果表明，所提的算法在Middlebury2005年数据集和MNIST数据集上的峰值信噪比分别达到32.97dB和28.26dB，相比于Pixel-RL算法分别提升了0.23dB和0.75dB，参数量和训练总时间降低了44.9%和18.2%，在边缘保持的同时有效降低了模型的复杂度。

Abstract: High-intensity Gaussian noise tends to blur or destroy the details and structure of the image, and result in the loss of edge information. Therefore, this paper proposes an edge-preserving image restoration algorithm based on pixel-by-pixel reinforcement learning. Firstly, a pixel-wise agent is constructed for each pixel. The algorithm takes the side window averaging filter at the edge into the action space. All pixel layer agents share the parameters of the advantage actor-critic algorithm, and the model can output the state transition probability of all positions at the same time and select the appropriate strategy for state transition to restore image. Secondly, the coordinated attention is combined in the feature extraction sharing network so as to focus the global information of all pixel positions between the feature channels and retain the information of position embedding. Then, to alleviate the problem of sparse reward, an auxiliary loss based on graph laplacian regularity is designed, which pays attention to the local smoothing information of the image and punishes the local unsmooth area so as to promote the pixel-layer agent to learn the correct strategy to maintain the edge more effectively. The experimental results show that the peak signal-to-noise ratio of the proposed algorithm on the Middlebury 2005 dataset and the MNIST dataset reaches 32.97dB and 28.26dB, respectively, which is improved by 0.23dB and 0.75dB compared with the Pixel-RL algorithm. The total parameter amount and training time are decreased by 44.9% and 18.2%, effectively reducing the complexity of the model while maintaining the edges.