Sonar And Radar Images Enhancement Based On Generative Adversarial Networks

Underwater sonar technology is significative for submarine object detection,but submarine sonar images are expensive and cannot be obtained on a large scale.Therefore,it is hard to process these images with deep learning algorithms.When the X-band nautical radar performs remote sensing on the ocean surface,the real-time ocean surface state can be acquired through the sea images and we can perform sea surface object detection.However,during the process of echo data acquisition,it will suffer a lot of noise problems and sea clutter interference,which will seriously affect the sea surface target detection.Considering the similarity of sonar and radar imaging principles,this paper studies the enhancement methods from super-resolution reconstruction and image denoising.This paper first uses the CycleGAN technology to make the underwater sonar images,but this technology is tough to produce high-resolution images.Whence a relatively lightweight super-resolution reconstruction network is proposed to enhance the manufactured images.For X-band nautical radar images denoising,inspired by(GAN-CNN Based Blind Denoiser)GCBD,this paper proposes a novel denoising strategy based on the Generative Adversarial Network(GAN).First,a GAN is trained to estimate the noise distribution over the X-band radar noisy images and to generate noise samples.Second,the noise patches sampled from the first step are utilized to construct a paired training dataset,which is used,in turn,to train our improved deep Convolutional Neural Network(CNN)for denoising.It is simple to operate and avoids complicated mathematical calculations.This paper is mainly divided into the following four research contents.:(1)Use CycleGAN to convert a large number of submarine optical images into sonar images.(2)A super-resolution network based on the Generative Adversarial Network(GAN)is proposed,which is added the improved Squeeze-and-Excitation(ISE).We improve the late excitation strategy,The test results of the benchmark show that the ISE structure can increase the peak signal-to-noise ratio(PSNR)of the image by about 0.1 dB.At the same time,the discriminator structure and the loss function are improved to output a multi-layer fusion feature map.Such an improvement slows down the problem of reducing reconstructionaccuracy after joining the adversarial loss.Finally,we use the super-resolution network to reconstruct low-resolution submarine sonar images..(3)The spectrally normalized GAN is used to model the real noise of the X-band rader images to expand the noise dataset,and then pair a large number of clean images to build the denoising dataset.We improve the denoising convolutional neural network(DNCNN)and add reversible downsampling layer and sub-pixel convolution layer to increase the perception field.Then,train the improved network(FDNCNN)on the constructed denoised dataset to realize the X-band radar image denoising.(4)Experiments verified the performance of the above methods.The experiment verified the feasibility of making low-resolution submarine sonar images by CycleGAN.The comparison results on the benchmark show that the proposed super-resolution network is superior to the current advanced super-resolution networks such as RCAN,RDN,EDSR and D-DBPN.It is verified by experiments that the spectrally normalized GAN can model the X-band radar image noise more realistically.In the experiment of verifying the influence of denoising datasets,we train FDNCCN on the spectrally normalized GAN enhancement dataset,the original noise dataset,and the full dark background dataset respectively.In the verification FDNCNN performance experiment,we train DNCNN,fast flexible convolution denoising network(FFDNet)and FDNCNN on the denoised dataset with the spectrally normalized GAN,The results show that FDNCNN can achieve the best denoising effect on the enhanced dataset.