基于残差学习的非对称卷积神经网络图像去噪方法  

Image Denoising Method Based on Residual Learning andAsymmetric Convolutional Neural Network

作　　者：曹阳 张英俊[1] 谢斌红[1] CAO Yang;ZHANG Yingjun;XIE Binhong(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024)

机构地区：[1]太原科技大学计算机科学与技术学院,太原030024

出　　处：《计算机与数字工程》2023年第6期1371-1375,1392,共6页Computer & Digital Engineering

基　　金：山西省重点研发计划重点项目(编号:201703D111027);山西省重点计划研发项目(编号:201803D121048,201803D121055)资助。

摘　　要：为了获得更清晰的图片,更好地去除图像中的噪声,对目前去噪效果较好的基于残差学习技术的深度卷积神经网络去噪算法(DnCNN)进行改进,提出了DnACNN模型,该模型通过将常规的方形卷积核替换为一组非对称卷积核来增强方形卷积核的骨架部分,从而提高特征提取精度和性能;并在测试阶段进行权值融合,从而减少参数、简化模型,确保不增加任何额外的推理时间。实验表明该方法在不增加任何额外推理时间的前提下能更有效地去除图像中的噪声,与目前主流去噪方法相比,能更好地捕获特征,提升特征丰富性,获得了更高的峰值信噪比,且能够更多地保留细节纹理。In order to get a clearer picture and better remove the noise in the image,this paper proposes DNA CNN model to improve the depth convolutional neural network denoising algorithm based on residual learning technology,which has good denoising effect at present.The model through the regular square convolution kernels is replaced by a set of asymmetric convolution kernels to enhance square frame part of the convolution kernel,so as to improve the feature extraction accuracy and performance.In the test phase,weight fusion is carried out to reduce parameters,simplify the model,and ensure that no additional reasoning time is added.Experiments show that this method can effectively remove the noise in the image without adding any extra reasoning time.Compared with the current mainstream denoising methods,it can better capture features,improve feature richness,obtain a higher PEAK SNR,retain more detail texture.

关 键 词：图像去噪 卷积核骨架 非对称卷积核 权值融合 特征提取 残差学习 参数 推理时间 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]