Research On Multi-frequency Image Noise Reduction Based On Deep Learning

Image denoising is an important problem in the field of digital image processing,which aims to remove noise from digital images to improve the quality and readability of images.Noise in digital images can be caused by various factors,such as errors in image acquisition equipment,interference in signal transmission,and errors in image processing.It is difficult to calculate the distribution of the noise caused by different noise inducing factors,and these noises will lead to random changes of pixel values or structural distortion of the image.Therefore,image noise will affect the visual quality of the image and the results of subsequent processing.In recent years,with the improvement of hardware equipment and the promotion and development of deep learning by researchers.Deep learning-based methods have made breakthroughs in the field of image denoising.However,the current image denoising algorithms based on deep learning mainly reduce the noise of images in the spatial domain.However,the noise mainly exists in the high frequency signal,so the high frequency signal is easily disturbed and difficult to recover.However,most of the current image denoising algorithms based on deep learning are carried out through deep learning.The noise distribution law in the noise image in the spatial domain is statistically analyzed,and the image denoising is carried out through the prior noise distribution law.However,in practical use,it is difficult to effectively reduce the noise in the high-frequency signal,which leads to the problem of poor image detail quality recovery.In order to solve the above problems and improve the overall image denoising performance,this thesis proposes an image denoising algorithm based on multi-band.The multiband based image denoising algorithm performs image denoising through three main parts.Firstly,the noise image was mapped from spatial domain to frequency domain by cosine transform at the input node of the noise image,and then the image in frequency domain was divided into high frequency signal and low frequency signal.Second,through the division of frequency domain signals,targeted noise reduction is carried out according to different frequency band characteristics.The low frequency signal mainly contains the structure information of the image,so the global image structure information is recovered in the low frequency signal.The noise reduction quality of low-frequency signals is improved by using multi-scale and other methods to improve the recovery of global image structure information.The high-frequency signal mainly contains the texture information such as the details of the image and the information in the high-frequency signal is sparse.Therefore,for high-frequency signals,on the basis of independent recovery of high-frequency signals,this thesis combines the characteristics of global information of low-frequency signals to assist the recovery of high-frequency signals through the attention mechanism.Third,after image denoising,the information of different frequency bands is fused.And signal enhancement is used to enhance and restore the weak details and texture information in the high-frequency signal.Through objective evaluation metrics,the multi-band based image denoising algorithm achieves advanced denoising performance on both real-world noise datasets and simulated noise datasets.And the effectiveness of the image denoising algorithm based on multi-band is proved by ablation experiments.