Research On Biological And Medical Image Denoising Method Based On Deep Learning

An important area of research in the field of computer vision is image denoising,which tries to use digital image processing methods to restore potentially clean images from degraded images tainted by noise.The denoising performance of models has been substantially enhanced by picture approaches based on deep learning thanks to the ongoing innovation and deepening of research.However,these techniques still have certain drawbacks,including a difficult time adapting to the content of biological images in real-world situations and poor performance in both synthetic noise and real-world noise environments of unknowable intensities.Deep learning-based techniques also rely on a lot of labeled data,which can be expensive to acquire and classify,particularly for biological imaging and medical imaging.It is challenging for the network to learn noise information when there aren’t any paired training data.This paper focuses on deep learning denoising algorithms applied to biological images and medical images acquired by electron microscopy and computed tomography(CT)imaging equipment.The following components comprise the main research efforts and accomplishments:(1)A double enhanced residual network based on image denoising is suggested to address the issue of synthetic noise on biological microscopic images.The suggested network,which has two subnetworks with the same structure,is a multilevel network with effective GPU memory utilization.To expand the network receptive field,features for each subnetwork were upscaled and downscaled through the hierarchical structure.The proposed method performs well in publicly available natural image datasets and then generalizes to biological microscopic images.The algorithm also produces pleasing visual effects for biological microscopic images with more complicated noise signals.(2)An adaptive projection network,a deep attention denoising network based on projection,is suggested as a solution to the issue of synthetic noise on medical CT images.This method,which is based on the U-shaped structure,separates noise and detailed texture using image adaptive projection and recovers the underlying clean image with texture features from the medical noise image.The proposed method outperforms other methods in terms of quantitative metrics and visual quality,according to experiments on medical CT images.(3)Aiming at the real noise problem on medical CT images,a blind denoising method for medical images based on noise generation network is proposed.For dental CT images without corresponding clean labels,a two-step framework is designed,including a noise generation network and a noise removal network.The noise generation network adopts the structure of generative adversarial network,which uses the learning ability of the generative model and the discriminative model to learn the real noise message,and then transfers to another set of clean CT images.New paired training data are constructed and used to train the noise removal network to complete denoising.This method can effectively remove the real noise,and the evaluation index is outstanding compared with the existing denoising methods.In this paper,the denoising task is designed for biological images and medical images.The data used include cell microscopic images,zebrafish microscopic images,low-dose lung CT images,and low-dose dental CT images.The involved image denoising models include Gaussian noise model,Gauss-Poisson noise model for biological microscopic images,and Poisson noise model,Gauss-Poisson noise model and real noise model for medical CT images.A large number of experiments show that the three algorithms proposed in this paper can generate clearer images,meet the image needs of the biomedical field,and have great application significance for practical scenes.