基于深度学习的合成孔径雷达图像去噪算法

Synthetic Aperture Radar Image Denoising Algorithm Based on Deep Learning

针对合成孔径雷达在生成图像时产生的固有斑点噪声问题,提出了一种基于多尺度注意力级联卷积神经网络的去噪算法,用多尺度卷积网络与注意力机制来实现图像的特征提取,将级联网络结构作为主网络中的特征增强部分,并加入批量归一化来防止模型中出现过拟合。实验结果表明,所提算法相较于其他传统图像去噪算法,峰值信噪比、结构相似性分别平均提高了0.75 dB~14.45 dB和0.01~0.16,在图像熵上也优于其他算法,并且能较好地恢复图像中的细节信息。

Objective Synthetic aperture radar(SAR)is a kind of sensor to capture microwave.Its principle is to establish the image through the reflection of waveform,so as to solve the problem that the traditional optical remote sensing radar is affected by the weather,air impurities and other environmental factors when collecting images.The most widely used SAR is its Change Detect(CD).Change detect is the dynamic acquisition of image information for a certain target,which includes three steps:image preprocessing,generation of difference map and analysis and calculation of difference map.It is applied to the estimation of natural disasters,the management and allocation of resources,and the measurement of land topographic characteristics.However,in the process of change detect,the inherent speckle noise in SAR images will reduce the performance of change detection.Therefore,image denoising method has become the basic method of preprocessing in change detection.How to restore a clean image from the noisy SAR image is an urgent problem to be solved.Methods The traditional denoising algorithm of SAR image generally uses the global denoising idea,and its principle is to use the global similar information in the image to process and judge.In the case of high resolution of the image,the algorithm needs a series of preprocessing such as smoothing,and then completes the pixel distinction through the neighborhood processing of each image block.The algorithm needs to occupy a large number of computing resources,and has certain spatial and temporal limitations in practical applications,and cannot efficiently complete the denoising task.In terms of deep learning,some algorithms perform well,but there is still room for improvement in network convergence speed,model redundancy and accuracy.To solve these problems,this paper proposes a denoising algorithm based on multi-scale attention cascade convolutional neural network(MALNet).The network mainly use the idea of multi-scale irregular convolution kernel and attention.Compared with a single convolution kernel,multi-scale irregular convolution kernel has a good image receptive field,that is,it can collect image information from different scales in order to extract more detailed image features.Subsequently,the convolution kernels of different scales are concat in the network,and the attention mechanism is introduced into the concat feature map to divide the attention of the features,so that the whole model has good enhancement ability for the main features of the image.In the middle of the network,the dense cascade layer is used to further strengthen the features,and finally the image restoration and reconstruction are realized by network subtraction.Results and Discussions In this paper,qualitative and quantitative experiments are carried out to evaluate and demonstrate the performance of the proposed MALNet model in denoising.The WNNM,SAR-BM3D and SAR-CNN algorithms are compared with our proposed methods.Visually observe the clear state and complete signs of the denoised image.In order to make a fair comparison,we use the default settings provided by the authors in the literature of the three algorithms to compare. Peak Signal to Noise Ratio ( PSNR ), Structural Similarity Index Measure( SSIM ) and image entropy are used as objective evaluation indexes.The PSNR, SSIM and imageentropy are calculated as error metrics.We use three denoising algorithms are compared, and aircraft, mountains and coasts areselected as verification images.The denoising effect of aircraft image (Fig.7), The denoising effectof coast image(Fig.8), and the denoising effect of mountains image (Fig.9). It shows the visual effectcomparison of denoising results of different algorithms when the current noise level is. In Figs.7, 8and 9, (a) is the noise-free image, (b) is the noise image, (c) is the denoising effect of the traditionalmethod WNNM, (d) is the denoising effect of SAR-BM3D, (e) is the denoising effect of SAR-CNN,and (f) is the denoising effect of MALNet proposed in this paper. It is obvious that the WNNMdenoising effect has many defects that are not removed clean, and the texture loss is quite serious.SAR-BM3D retains some details, but the aircraft fuselage is very vague, the tail part has been erasedmost of the edge information. Although the aircraft wing in the SAR-CNN denoising effect imageis recovered, the whole aircraft at the bottom is still far from the reference image, and the recoveredsmall objects are blurred.It can be seen from (Table 3) that the average PSNR value of the proposed MALNet is about9.25 dB higher than that of SAR-BM3D, about 0.75 dB higher than that of SAR-CNN, and about14.45 dB higher than that of WNNM. Moreover, in addition to the noise level, MALNet is 0.01 dBless than that of SAR-CNN. The PSNR value of the proposed MALNet model at each noise level ishigher than that of the comparison of the three values. Especially when the noise parameter is 20,the proposed method is 2.56 dB higher than that of the comparison of SAR-CNN algorithm. In termsof structural similarity, it can be seen that the structural similarity of MALNet is mostly the highestvalue in the comparison method. Only when the noise parameter is 50, it is slightly lower than thatof SAR-CNN, and the average structural similarity is also the highest. The average informationentropy of the denoised images by the four algorithms is 7.113492 for WNNM, 6.842258 for SARBM3D,7.499375 for SAR-CNN and 6.6917 for MALNet. The proposed algorithm outperformsWNNM, SAR-BM3D and SAR-CNN by 0.42179, 0.15056 and 0.80768, respectively. Therefore,considering the three objective evaluation indexes of PSNR, SSIM and image entropy, the proposednetwork in this paper has better denoising performance than the comparison method.Conclusions In this paper, a new denoising model MALNet is proposed for the noise in SAR images.This model uses the end-to-end architecture and does not require separate subnets or manualintervention. The solution includes three modules : multi-scale irregular convolution module,feature extraction module of attention mechanism and feature enhancement module of densecascade network. The model also adds batch normalization and global average pooling to improvethe adaptability of the model. It can complete the convergence without a large number of data sets.The image data complete the convergence after 150 rounds of training. The training efficiency isoutstanding and the portability is good.The experimental results show that, compared with othertraditional image denoising algorithms, the peak signal-to-noise ratio and structural similarity of theproposed algorithm are improved by 0.75 dB ~ 14.45 dB and 0.01 ~ 0.16, respectively. The proposedalgorithm is superior to other algorithms in image entropy, and can better recover the details of theimage.