Image noise reduction by deep learning methods

ABSTRACT


INTRODUCTION
There are various machine learning methods for removing noise in images, including neural networks, deep learning, image recovery methods [1], and regularization methods [2]- [4].Neural networks [5] and deep learning, such as autoencoding and convolutional neural networks (CNNs) [6], [7] are often used to remove noise in images.In this paper, the methods considered were trained on clean images, and then used this model to remove noise on noisy images.Raindrops, traces of snow, and dust in the images can be considered as one of the types of noise in the images.They are random white or black spots, and white or gray lines on the background of the image, which can degrade the image quality and make it difficult to analyze.When shooting outdoors, these noises can be especially noticeable on dark backgrounds and at night.To remove such noise in images, various machine learning methods can be used [8]- [10] such as filters, CNNs, and other algorithms.CNNs can be trained on a large number of images with different types of noise to automatically remove noise on new images.There are also noise reduction methods based on statistical analysis that can help in removing such types of noise in images.For example, the histogram method is used to improve image quality by applying histogram alignment.This method changes the brightness or color characteristic of the image in such a way as to obtain a more uniform distribution of pixel values.This method can help to improve the image quality and increase its readability.
Lyu et al. [11] present a fast and flexible denoising network (FFDNet), a deep learning model that focuses on fast and flexible image noise reduction.The model provides competitive results in the field of noise reduction while maintaining efficient computing performance.According to Ulyanov et al. [12] the concept of using deep neural networks as prerequisites for image reconstruction tasks, including noise suppression, is investigated.The authors demonstrate that a randomly initialized deep neural network (DNN) can capture the basic structure of images and effectively eliminate their noise without the need for additional training data.
Batson and Royer [13] present Noise2Self, a self-supervised learning method for denoising images.The approach leverages the idea that an image can be used as it is noisy label, allowing the model to learn denoising directly from noisy images.Zhai et al. [14] addresses the challenges of real-world image denoising, where the noise characteristics are unknown.The authors propose neural adaptation networks (NANs), which can adapt to different noise levels and distributions for effective denoising.
Zhang et al. [15] focuses on learning a deep CNN denoiser prior, which can be applied to various image restoration tasks, including denoising.The authors propose a two-stage training framework that incorporates both noisy and clean images to learn effective denoising prior.Sharan et al. [16] focuses on video denoising, it presents a deep learning-based approach that can be extended to image denoising.The authors propose a deep video prior, which exploits temporal redundancy in videos to effectively denoise individual frames.These are just a few notable works that demonstrate the effectiveness of deep learning techniques to reduce noise in images.It is important to note that research in this area is constantly evolving, so conducting a comprehensive literature search would allow you to get a more detailed idea of the achievements made in this area.

METHOD
The generative model of the attentive generative adversarial network (GAN) combines the ideas of attention and classical GAN architecture.The AttentiveGAN [17], [18] allows the model to focus on specific areas in the data and generate better results.Attention in the context of machine learning refers to a mechanism that allows the model to focus on certain parts of the input data, ignoring the rest.This allows the model to focus on more important details or data areas.An AttentiveGAN usually consists of two main components: a generator and a discriminator, which compete with each other in the learning process.The generator takes random noise or another form of input data as input and generates the generated data.Attention can be applied to the generator so that it can focus on certain areas of data and create more realistic results.The discriminator accepts both real data and generated data from the generator and tries to distinguish them.The task of the discriminator is to distinguish real data from fake data with maximum accuracy.Learning occurs by opposing the generator and discriminator to each other.The architecture of the model for the applied task of this work is shown in Figure 1.[20]- [22].After passing through the AGAN model, the resulting cleaned image is further processed using Gaussian filtering [23]- [25].This allows you to smooth out the remaining noise and improve the quality of the final image.

RESULT AND DISCUSSION
To solve this problem, noisy images with different noise densities were taken, which contain raindrops and traces of dust and snow.The training data set contains 15,314 pre-trained sets of images taken from the Kaggle open-access database.Figure 2 shows noisy images with various types of noise, such as Figure 2 In this work, deep learning methods show high efficiency in reducing noise in images.They allow models to automatically study complex dependencies in data and restore clean images with high accuracy.The Terrain model was used for all training images, which made it possible to reduce noise in images of different types.Figure 4 shows the algorithm for executing the AttentiveGAN model using a Gaussian mask.Gloss allows you to estimate the difference between the generated data and the real data.The lower the Gloss, the better the generator performs its task.
The discriminative loss (D-loss) loss measure evaluates how well the discriminator in GAN distinguishes between generated data and real data.The discriminator in GAN is a model that is trained to classify data as real or generated.allows you to estimate the difference between the discriminator classification and the actual data labels.The purpose of the discriminator is to be able to accurately distinguish real data from generated data.The lower the D-loss, the better the discriminator performs it is task.The structural similarity index (SSIM) metric is used to measure the similarity between two images.SSIM evaluates the structural similarity between the original and restored images.It takes into account the brightness, contrast, structure, and perception of human vision.SSIM outputs a value from 0 to 1, where 1 means perfect similarity.A high SSIM value as shown in Figure 5 indicates that the restored image is close to the original one in terms of structure.

CONCLUSION
In this experiment, the resulting improved images can be effectively used in the future in object recognition.If the background in an image contains noise or unwanted detail, blurring the background helps make objects stand out more.Automated image noise reduction using deep learning techniques can make objects more distinct and easier to recognize.
In the course of the work, it was noted that the effectiveness of deep learning methods for noise reduction depends on the quality of training data, the amount of data, and the choice of model architecture.In this work, noisy images with different noise densities were taken.The optimal result can be achieved with the correct selection of parameters and optimization of the model for a specific type of noise and the desired results.In this work, to achieve a better result, we used a combined model, that is, we trained the data using the Derain model and applied a Gaussian filter.To assess the effectiveness of the model, Gloss, D-loss, and SIM indicators were used, where the SIM value reaches 1, which allows us to evaluate the effectiveness of the deep learning method used.

Figure 1 .
Figure 1.The architecture of the AttentiveGAN model

Figure 4 .
Figure 4. Algorithm for executing the AttentiveGAN model