基于改进机器学习的超分辨率图像细节复原

Super-Resolution Image Detail Recovery Based on Improved Machine Learning

超分辨率图像细节复原技术中存在图像纹细节恢复效果差等问题。针对上述问题,提出一种基于改进机器学习的超分辨率图像细节复原算法。算法首先融合空间注意力机制并采用1个卷积层与1个激活层对图像的纹理特征进行提取;然后融合通道注意力机制利用3个深度可分离卷积层提取图像的语义特征,提高特征提取效率;接着对提取的双特征进行串接融合成一组特征图,并采用上采样的方法通过对特征图学习,将HR图像重建为SR图像;最后采用局部细节增强模型对重构的SR图像的纹理细节进行二次增强,提高图像的信息表达能力。仿真结果表明,上述算法在PSNR和SSIM两个评价指标上较主流算法均有明显提升;在Urban100数据集上,在不同重构倍数下上述算法均取得了最高的评价指标值,较传统算法其PSNR和SSIM分别平均提高1.303dB与0.0293dB,具有较好的鲁棒性。因此提出的基于改进机器学习超分辨率图像复原算法可有效重建出纹理细节更丰富的超分辨率图,高效率高质量的恢复图像纹理细节信息。

There are problems in super-resolution image detail recovery techniques,such as poor image grain de⁃tail recovery.To address these problems,this paper proposes a super-resolution image detail recovery algorithm based on improved machine learning.The algorithm firstly fuses the spatial attention mechanism and uses one convo⁃lutional layer and one activation layer to extract the texture features of the image;Then the channel attention mecha⁃nism is fused to extract the semantic features of the image using three depth-separable convolutional layers to improve the feature extraction efficiency;Next,the extracted double features are concatenated and fused into a set of feature maps,and the up-sampling method is used to reconstruct the HR image into the SR image by learning the feature maps;Finally,the reconstructed image is reconstructed into the SR image using the local detail enhancement model into SR image,and the local detail enhancement model is used to enhance the texture details of the reconstructed SR image twice to improve the information expression ability of the image.The simulation results show that the algorithm has significantly improved the evaluation indexes like PSNR and SSIM compared with the mainstream algorithms;on Urban100 dataset,the algorithm achieves the highest evaluation index value with different reconstruction multiples,and the PSNR and SSIM are improved by 1.303dB and 0.0293dB respectively on average compared with the tradi⁃tional algorithm,which has better robustness.Therefore,the improved machine learning-based super-resolution image recovery algorithm proposed in this paper can effectively reconstruct super-resolution maps with richer texturedetails and recover image texture detail information with high efficiency and high quality.