基于深度学习和稀疏编码的图像超分辨率重建

IMAGE SUPER-RESOLUTION RECONSTRUCTION BASED ON DEEP LEARNING AND SPARSE CODING

针对基于深度神经网络的图像超分辨率重建技术训练时间长的问题,提出一种基于深度学习和稀疏编码的图像超分辨率重建算法。采用卷积神经网络学习低分辨率图像每一块的深度视觉特征,利用局部约束线性编码的局部平滑稀疏能力对深度特征进行编码;利用字典学习技术学习低分辨率图像和高分辨率图像每一块之间的判别关系字典;通过低分辨率字典和低分辨率图像估计稀疏表示系数,利用该系数实现图像超分辨率的重建。实验结果表明,该算法在视觉效果和评价指标上均获得了较好的超分辨率效果,并且速度较快。

Aiming at the problem that the image super-resolution reconstruction techniques based on deep neural networks need long training time, we propose an image super-resolution reconstruction algorithm based on deep learning and sparse coding. Convolutional neural network was adopted to learn deep visual features of each block of low resolution images. The local smoothing and sparsity ability of locally constrained linear coding was used to encode depth features. We applied the dictionary learning technique in learning the discriminant relationship dictionary between each block of low resolution image and high resolution image. The sparse representation coefficient was estimated through low resolution dictionary and low resolution image. The coefficient was used to reconstruct the image super resolution. Experimental results indicate that the proposed algorithm realizes good super resolution effect on both visual quality and evaluation indices, and it also has fast speed.