自适应分数阶微分图像超分辨率算法研究  

Research on Adaptive Fractional Order Differential Image Super-resolution Algorithm

作　　者：王仕达 杨旗[1] 张萌 Wang Shida;Yang Qi;Zhang Meng

机构地区：[1]沈阳理工大学机械工程学院,辽宁沈阳110159 [2]北华大学,吉林吉林132013

出　　处：《一重技术》2024年第5期50-55,49,共7页CFHI Technology

摘　　要：超分辨率技术是在低分辨率图像中估计高分辨率图像的一种应用。尽管深度神经网络在这方面已经取得显著进展,但仍存在一些问题,如对微小结构变化的敏感性。对此,通过引入残差密集网络提升GAN网络的灵活性;此外,针对分数阶微积分在神经网络领域的应用,提出FoMCA分数阶多通道注意力机制模块,通过设计新的阶数调节函数提高模型性能和效率,为构建和训练非常深层的可训练网络提供新思路;最后,通过对比实验和消融实验验证上述方法的有效性。Super-resolution technique is an application of estimating high-resolution images from low-resolution images.Although significant progress in this field has been achieved by deep neural network,there are still some problems,such as sensitivity to small structural changes.Therefore,residual dense network is applied to improve the flexibility of GAN network;in addition,for the application of fractional order calculus in the field of neural network,the FoMCA fractional order multichannel attention mechanism module is proposed,which improves the performance and efficiency of the model through the design of new order adjustment function,which provides a new way of thinking for constructing and training of deeper trainable network;finally,comparison experiment and ablation experiment are carried out to verify effectiveness of the above methods.

关 键 词：分数阶微积分 注意力机制 残差密集网络 神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]