摘要
为了更有效地去除图像中存在的高斯噪声,提出一种结合Inception模块的卷积自编码器图像去噪模型。以完整图像作为输入和输出,利用Inception模块对噪声图像进行去噪,使用改进Inception反卷积模块将去噪图像进行还原,提升模型去噪能力。同时在模型中引入批量归一化(Batch Normalization,BN)和随机失活层(Dropout)有效解决过拟合问题,引入ReLU函数避免模型梯度消失,加速网络训练。实验结果表明,与深度卷积神经网络方法相比,该模型获得了更高的峰值信噪比和结构相似度,其去噪能力更好,视觉效果更佳,具有更好的鲁棒性。
This paper proposes a convolutional auto-encoder image denoising model combined with inception module in order to remove the Ggauss noise in the image more effectively.Take the whole image as the input and output,it used Inception module to denoise the noise image,and the improved Inception deconvolution module was used to restore the denoised image to improve the denoising ability of the model.We also introduced BN and Dropout to prevent over-fitting and used ReLU function to avoid gradient disappearance and speed up network training.The experiment results show that compared with the deep convolution neural network method,the proposed method can remove the Gauss noise more effectively,obtain higher peak signal-to-noise ratio and structural similarity,and has better visual effect and robustness.
作者
雷景生
闫晨阳
杨忠光
Lei Jingsheng;Yan Chenyang;Yang Zhongguang(College of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200082,China)
出处
《计算机应用与软件》
北大核心
2021年第2期221-226,322,共7页
Computer Applications and Software
基金
国家自然科学基金项目(61672337)。
