Research On Image Denoising Technology Based On N2N And Attention Mechanism

Images are one of the direct sources of information in people’s daily lives,widely used in fields such as weather forecasting,astronomy,and medical diagnosis.However,during the imaging process,certain inherent factors(such as sensors and imaging devices)and external factors(indoor/outdoor environment)can interfere with pixel generation,resulting in blurred images.On the one hand,such images have poor visibility,hindering people from directly obtaining information from them.On the other hand,the error information obtained from noisy images is highly likely to lead to research errors in special fields,resulting in irreversible harm.Thus,image denoising technology has emerged,and many scholars have proposed various methods for image denoising,including traditional methods and deep learning methods.Traditional methods use a large number of filtering algorithms to restore image accuracy,but they have high computational complexity and low efficiency.Deep learning-based methods use a large number of clean image pairs to train neural networks to learn the ability to restore clear images,addressing the pain points of traditional methods,but also facing some challenges:(1)scarcity of clean image training samples;(2)existing deep denoising networks are prone to losing edge details of image features;(3)poor generalization ability of models.This thesis aims to solve the above challenges and attempt to break through the bottleneck of deep learning methods in practical application scenarios.Specifically,this thesis conducted the following research:(1)To address the problem that deep neural networks extract image features that are prone to neglecting detail information,a multi-scale residual convolutional block network with attention mechanism was proposed,and noise-noise(Noise2Noise,N2N)image pairs were used to train the neural network.The multi-scale residual blocks of the network can excavate critical feature information of images at different network depths and effectively transmit it to deeper layers of the network.Its residual path directly transfers shallow features to deep networks,which fused to obtain richer feature maps,making the trained denoising network capable of restoring clearer images.Comparative experiments with two models in N2 N and other classical methods showed that the model restored image details more clearly,and its objective metric data was also better.(2)To improve the denoising network’s control over global information and generalization ability,a multi-head attention mechanism combined with a simple residual block image denoising network was further proposed.The multi-head attention mechanism can help the network master more global feature information and the relationship between feature information during the learning process;the simple residual block uses residual connections to reduce model complexity and enhance the network model’s generalization ability.The final trained model can perform multiple image denoising tasks,including medical MRI undersampling reconstruction.Comparative training with two models in N2 N showed that the model converged faster and more stably in the same range of noise amplitudes.Compared with classical denoising methods,the model performed better on evaluation metrics.At the same time,the model performed well in MRI undersampling reconstruction,verifying its strong generalization ability.