深度融合多尺度结构信息的图像超分辨率算法

A multi-scale structure information fusion based image super resolution algorithm

现有基于深度学习的图像超分辨率算法主要通过增加卷积神经网络的深度获得良好的图像重建精度,但这些方法缺少对低分辨图像中纹理信息的多尺度分析,以至一些重要的细节特征在超分辨率图像重建中很难得到有效保持。提出了一种基于多尺度结构信息堆叠融合的图像超分辨率算法,该方法通过堆叠多个结构信息金字塔模块,实现对低分辨率图像由浅到深、从局部到全局的特征提取和融合,其中的结构金字塔模块由密集连接注意力编码器和逐级反卷积解码器构成,可以充分提取并融合多个局部尺度的结构信息。在Set5、Set14、B100和Urban100上与现有算法进行比较,实验结果表明:所提方法在评价指标和视觉上效果更优,尤其在复杂场景、较大倍数、含噪的情况下效果更好。

The existing deep learning based on image super-resolution algorithms mainly obtain good image reconstruction accuracy by increasing the depth of convolutional neural network,but these methods are lack of multi-scale analysis of texture information in low-resolution images,so that some important detailed features are difficult to be effectively maintained in super-resolution image reconstruction.In the paper,an image super-resolution algorithm based on multi-scale structure information stacking fusion is proposed.The method through the stack multiple pyramid structure information module,realizes the low resolution image multi-scale feature from local to global effective feature extraction and fusion,in which the structure of the pyramid modules are connected by a dense deconvolution decoder attention encoder and step by step,can be fully extracted and texture structure information fusion of multiple local scale.The method in this paper is compared with the existing algorithms on Set5,Set14,B100 and Urban100.The experimental results show that the proposed method has better evaluation indexes and visual effects,especially in the case of complex scenes,large multiples and noise.