基于卷积神经网络的暗光图像去噪算法研究

Research on dark-light image denoising algorithm based on CNN

针对暗光图像往往存在大量分布不均的噪声,极大地影响图像质量,而现有的基于单阶段卷积神经网络(CNN)的模型不能有效去除大量暗区域噪声的问题,提出一种基于CNN的暗光图像去噪算法模型。通过两种尺度的特征映射去噪模块共同构成深层CNN模型,合理运用残差学习与类似自编码器单元有效地重构出去噪图像;采用结构相似性(SSIM)作为损失函数训练模型。实验结果表明:预训练模型在BSD68数据集的峰值信噪比(PSNR)和SSIM值可同时达到25.23 dB和0.927,对自然场景的噪声图像恢复的PSNR和SSIM达到14.03 dB和0.423。本文模型对高斯白噪声和暗光条件的去噪效果显著,对自然暗光场景图像有着较好的对比度恢复和去噪效果。

Aiming at the problem that dark-light images often have a large amount of unevenly distributed noise,which greatly affects image quality,however,the existing models based on single-stage convolutional neural network(CNN)cannot effectively remove a large amount of dark area noise,a CNN-based dark-light image denoising algorithm model is proposed.Two scales of feature mapping denoising modules are used to form a deep CNN model,and the residual learning and similar autoencoder units are used to effectively reconstruct the denoised image.The structural similarity(SSIM)is used as the loss function to train the model.The experimental results show that the peak signal-to-noise ratio(PSNR)and SSIM value of the pre-trained model in the BSD68 dataset can reach 25.23 dB and 0.927 at the same time,and the PSNR and SSIM for the restoration of noisy images of natural scenes can reach 14.03 dB and 0.423.The model in this paper has a significant denoising effect on Gaussian white noise and dark light conditions,and has a better contrast restoration and denoising effect on natural dark light scene images.