基于双残差注意力网络的ICMOS图像去噪算法

Dual Residual Attention Network for ICMOS Sensing Image

针对微光夜视条件下增强型CMOS(ICMOS)图像信噪比低、随机噪声明显的问题,提出了一种基于双残差注意力网络的ICMOS图像去噪算法。为了制作ICMOS噪声图像数据集,采用ICMOS相机在特定照度环境下拍摄静态噪声图像序列,然后采用多帧平均的方法获得每个序列对应的无噪声真值图像;其次,为了直接从噪声图像中提取噪声分量,设计了一种结合噪声残差学习和残差网络模块的双残差网络模型,并引入通道注意力机制给模型的特征图维度赋予权重,在提升模型学习能力的同时降低了模型复杂度;最后,采用网络训练所得模型对测试图像进行去噪实验。对比实验结果表明,本文提出的算法得到的峰值信噪比较经典的BM3D算法提升了9.56 dB,结构相似度提升了0.0503。从主观效果可以看出,本文算法可以更好地去除ICMOS图像噪声,保留图像细节,同时,具有较高的运行效率。

Low-light-level night vision technology is to explore the photoelectric technology that how to enhance,transmit,store,reproduce and apply the images captured under low light conditions.It is an important part of modern optoelectronic technology.ICCD/ICMOS(Intensified CCD/CMOS)is a solid low-light imaging device with a wide range of applications and the lowest working illuminance which is formed by coupling an image intensifier and CCD/CMOS.Although ICMOS can image under low-light night vision conditions,the image intensifier also amplifies the intensity of the noise while enhancing the signal,resulting in obvious random noise in the captured image,and the noise characteristics are more complex than that of traditional CMOS imaging.Due to the microchannel plates,ICMOS sensing image noise is not independent and identically distributed,but aggregated random noise with spatial correlation.Aggregated noise destroys the original structural features of the image,which also greatly increases the difficulty of denoising.In this paper,we propose a dual residual attention network for ICMOS sensing image denoising.There are three main ideas for our method.First,the network adopts the idea of residual learning,which means that the output of the network is the noise image,not the denoised image.Then the denoised image is achieved by subtracting the noise image from the original image.The residual learning network only needs to extract the noise component from the original image,which greatly reduces the difficulty of training the network.Secondly,we introduce four residual attention modules in our model,and the number of feature maps of each module is constantly decreasing.Each residual attention module consists of four residual blocks,one channel attention layer and one convolutional layer.The basic unit of the module is the residual block,which can effectively improve the network performance.At the same time,the introduction of the residual module can better solve the problems of gradient dispersion,gradient explosion and gradient degradation.Finally,the network introduces the channel attention layer,which can assign different weights to the output feature map of the middle layer,thereby analyzing the importance of each feature channel,and then enhancing the useful features and suppressing slight features according to this importance,and finally guide the network to continuously reduce the dimension of the feature map.Existing deep learning denoising methods mostly work for simulated Gauss-Poisson distributed noise and real noise data of some natural images.These methods can not be directly applied to ICMOS sensing images.Due to the particularity of ICMOS imaging noise,we made the ICMOS image dataset ourselves.We adopt the multi-frame averaging method to obtain the label image The image sequence is captured from a static scene under a certain fixed illumination in the dark room,and then one label clean image of the image sequence is synthesized by a multi-frame weighted average method.The scene illuminance is accurately measured with an illuminance meter.This dataset is mainly based on three different illuminances 2×10^(−1)、3×10^(−2)、2×10^(−3) lx for image acquisition,and seven different static scenes are collected under each illuminance condition.Due to the inconsistency of noise intensity and brightness,we conduct model training for images under different illuminances.Two static scenes with 1000 images are used as training sets under each illuminance.Our method applied the L1 loss as the loss function.From the subjective and objective results,it can be seen that our method has better denoising results and higher efficiency than other state-of-art methods.