Algorithm for improving the quality of mixed noisy images

Digital images are one of the effective means of receiving, transmitting, exchanging and expressing information. However, the digital image is not always of the expected quality, that is, as a result of the influence of various factors, noises appear in the image. This significantly reduces the accuracy and quality of the data in the image. Segmentation and recognition of objects from a low-quality image creates many additional problems. For example, it requires image quality enhancement based on powerful pre-processing algorithms. One such problem is to eliminate multiple noises in the image, and it is one of the pressing issues of digital image processing. This paper is devoted to solving the problem of image reconstruction by eliminating the mixed condition of Gaussian, salt-pepper and Poisson noises, which are common in digital images, in which one filter that optimally reduces each type of noise is selected and all combinations of them are applied to the mixed noise image to maximize the images. The idea of restoration has been put forward. The quality of the reconstructed images was evaluated by the NIQE criterion, which is one of the no-reference quality indicators, and a rule for automating the image processing process was proposed based on its values.


Introduction
Today, digital images are one of the main tools in the process of receiving, processing and exchanging information.Digital images have become a powerful tool for communicating information, ideas, feelings and experiences on social media, email or online platforms around the world.Also, the importance of medical digital images in the field of health is great, and it helps to predict and treat existing diseases in a person.
Many field problems can be solved by recognizing and classifying objects in digital images [1].In the field of computer vision, one of the important stages that serve to increase the accuracy of recognition of objects in the image is the stage of pre-processing of images.Because the results of this stage do not affect the results of the next stages [2][3][4][5][6][7].The pre-processing step provides the opportunity to obtain a quality image for the input of the next step.
One of the main steps in image preprocessing is image noise removal.The presence of noise in the image causes a significant decrease in image quality.Noise is an unwanted element added to the original clean image during image acquisition or transmission.In the real world, an image can be affected by several types of noise at the same time.In this paper, mixed noise images with simultaneous addition of Gaussian, salt-pepper and Poisson types of noise, which are common in digital images, are studied.High temperature in the imaging process, insufficient light causes Gaussian noise in the image [8].Salt-and-pepper noise is generated in the image during image digitization or as a result of incorrect memory allocation [9].Poisson noise depends on the statistical nature of electromagnetic waves, such as X-rays and gamma rays, and is caused by random photons [10].In some cases, these three noises can meet together, that is, in a mixed form.Until now, the existing noise reduction approaches are mainly designed to eliminate one type of noise in the image, that is, only one type of noise is added to the original clean image.Considering that there are many effective algorithms for removing single type of noise, this paper develops an algorithm based on a new approach to improve image quality by removing mixed noise.
In evaluating the effectiveness of the proposed algorithm, the NIQE criterion, which is calculated as a quality indicator without a reference, was selected, and an image processing rule was developed based on its values.The authors believe that the use of the proposed rule contributes to the increase in the speed of recognition of objects in the image.

Methods
The optimal filter for each type of noise was selected based on a comprehensive analysis of the literature, that is, BM3D [11] for G n , median [12] for S n , TV filters were selected for P n [13][14][15].The application of these filters to an image is shown in the following program code: 1) def apply_bm3d(noisy): denoise = bm3d.bm3d(noisy,sigma_psd=0. The image resulting from the application of (2) to the image (1) By evaluating the quality of the generated image based on (3), one can evaluate how close it is to the original image.In this case, the NIQE indicator, which evaluates the image quality without a reference, was chosen as the B operator.
Condition ( 4) is used to determine the optimal filter among hybrid filters.A small value of NIQE determines the efficiency of the filter.

Computing experience
The research used 52 image samples from the database of images cited on www.kaggle.com.Noise was added to these images in various combinations to create mixed noise images (figure 1).(2) filters were tested on each image according to (4) conditions, and finally, the hybrid filter that met the most (4) conditions based on the total number of images was selected as the optimal filter for the noise sequence (table 1).

Figure 1 .
A sample of images created from a combination of noises.Applying a sequence of matching filters to the noise sequence added to the original image produces (2) filters, and applying them to each of the images shows a sample of the resulting images in the figure below.

Figure 2 .
Figure 2. Examples of images resulting from the application of (2) filters to an Poisson noise.The output of generated noise along with parameter values is given in Initially, different noisy images are generated from the combination of G n − Gaussian, S n − salt- pepper and P n − in terms of combinations is denoted Î .

Table 1
shows that the optimal filter sequence for all cases of noise sequences is