多注意力机制引导的双目图像超分辨率重建算法

Binocular image super-resolution reconstruction algorithm guided by multi-attention mechanism

由于水下环境复杂,采集的水下图像通常是退化的低质图像。因此提出一种多注意力机制引导的双目图像超分辨率重建算法,选择性挖掘学习图像特征信息,实现高质量图像重建。针对水下图像分辨率低问题,引入双层注意力机制来加强重要细节特征的学习;然后针对双目图像的视差特性,提出一种视差注意力机制来充分学习左右目图像的先验信息,有效提高了图像质量。在Middlebury数据集2倍和4倍重建图像的信噪比分别为33.3和28.39 dB,表明该算法可以在提高图像空间分辨率的同时保留图像细节信息;同时该算法在拍摄的真实水下图像上的重建效果优于其他算法,表明其能实现更高质量的水下图像超分辨率重建。

Due to the complex underwater environment,underwater images are usually degraded low-quality images.Therefore,a multi attention mechanism guided binocular image super-resolution reconstruction algorithm is proposed to selectively learn image feature information for achieving high-quality image reconstruction.Aiming at the low resolution of underwater image,a network with double attention module is designed to enhance the learning of important details.Then,aiming at the disparity characteristics of binocular images,a parallax attention module is proposed to fully learn the prior information of left and right-hand images,and improve the image quality effectively.The PSNR of the reconstructed image with×2 and×4 on the Middlebury dataset is 33.3 and 28.39 dB respectively.It shows that the algorithm can improve the spatial resolution of the image and better retain the image details.At the same time,the reconstruction effect of this algorithm is better than other algorithms on the underwater dataset in real underwater scenes,indicating that it can achieve higher quality underwater image super-resolution reconstruction.