结合非对称卷积与特征蒸馏的图像超分辨率重建网络

Image super-resolution reconstruction network combining asymmetric convolution and feature distillation

为了进一步提高单幅图像超分辨率(single image super-resolution,SISR)轻量化网络的图像重建效果,基于轻量化网络RFDN,提出一种结合非对称卷积与特征蒸馏的图像超分辨率重建网络(asymmetric convolution distillation network,ACDN)。首先利用非对称卷积构建特征提取模块,在残差块中并联2个不同卷积核的非对称卷积,增强网络对特征的提取能力;其次利用均衡注意力机制与非对称卷积改进特征蒸馏模块,强化网络对高频信息的获取;最后在重建模块中加入均衡注意力机制进一步提高网络的最终重建性能。实验结果表明:与RLFN、SMSR等先进轻量化网络相比,提出的ACDN网络能在5个标准数据集上重建出纹理细节更丰富的高质量图像,重建图像的峰值信噪比和结构相似性指标均有提升,并在网络模型的参数量和性能上达到了更好的平衡。

In order to further improve the image reconstruction effect of single image super-resolution(SISR)lightweight network,based on lightweight network RFDN,an image super-resolution reconstruction network combining asymmetric convolution and feature distillation was proposed.Firstly,asymmetric convolution was used to construct a feature extraction module,the asymmetric convolution of two different convolution kernels in parallel in the residual block enhances the feature extraction capability of the network.Secondly,the balanced attention mechanism(BAM)and asymmetric convolution were used to improve the feature distillation module for the acquisition of high frequency information.Finally,BAM was added to the reconstruction module to further improve the final reconstruction performance of the network.The experimental results show that compared with advanced lightweight networks such as RLFN and SMSR,the proposed ACDN can reconstruct high-quality images with richer texture details on five standard data sets,improve the peak signal-to-noise ratio and structural similarity index of reconstructed images,and achieve a better balance between the number of parameters and the performance of the network model.