面向盲去模糊的选择性内核卷积混合连接编解码网络

针对运动模糊在空间上非均匀且模糊核未知的特点,提出一种选择性内核卷积混合连接编解码网络.为避免估计模糊核产生误差,该网络采用对称编解码的结构,以端到端的方式实现盲去模糊;为消除非均匀模糊,设计一种选择性内核卷积混合连接块,使用两个感受野不同的分支增强网络提高非均匀模糊特征的适应能力;同时,该模块融合了通道注意力机制,对多分支特征图进行校准,增强有效特征信息;此外,使用均方误差与感知损失作为联合损失函数引导网络模型的训练.实验结果表明,该网络能够有效去除图像模糊,恢复出图像的边缘结构和纹理细节.

Selective-kernel convolution mixed link encoder-decoder network for blind deblur

Aiming at the characteristics that motion blur is spatially non-uniform and the blur kernel is unknown, a selective-kernel convolution mixed link encoder-decoder network was propose. In order to avoid errors of blur kernel estimation, a symmetrical codec structure was applied by this network to complete blind deblurring. In order to remove non-uniform blur, a selective-kernel convolution mixed link block was proposed, which used two branches with different receptive fields to enhance the adaptability of the network to non-uniform blur features; At the same time, channel attention mechanism was integrated with the block to calibrate feature maps from different branches and enhance useful channel information; In addition, Mean Square Error and Perceptual Loss are used as a joint loss function to train the network model. Experiment results show that the proposed network can effectively deblur image while preserving the edge structure and texture details of the image.