Recursive reasoning image restoration method based on attention mechanism

The invention discloses a recursive reasoning image restoration method based on an attention mechanism. The method comprises the following steps: acquiring a to-be-processed image Y and a mask image; combining the to-be-processed image Y and the mask image, and obtaining a feature map X with a fixed number of channels in a covered area through a local convolution operation with stride being 2; processing the feature image X and the mask image by using a recursive reasoning module to obtain a feature image set F and a mask image set M; combining all the feature maps in the feature map set F by adopting an adaptive function to obtain a new feature map R; and performing up-sampling operation on the feature map R to obtain a repaired image. According to the invention, accurate and efficient image restoration can be realized. 本发明公开了一种基于注意力机制的递归推理图像修复方法，包括获取待处理图像Y和mask图像；将待处理图像Y和mask图像进行组合，并通过stride为2的局部卷积操作得到存在遮掩区域固定通道数的特征图X；使用递归推理模块对特征图X和mask图像进行处理，得到特征图集合F和遮掩图集合M；采用自适应函数合并特征图集合F中的所有特征图，得到新的特征图R；对特征图R进行上采样操作，得到修复