With the increasing popularity of digital cameras, people will encounter various images in their daily lives. Some images are of good quality, while some of the images we encountered are of poor quality. Due to the presence of noise, the image quality will also decrease. This noise may be caused by low light conditions or other intensity issues. In order to denoise the image, that is, reduce the noise in the image, there are many methods that can be used. It has been a hot topic of research for a long time, and it is still being experimented by researchers. Here, we will discuss how to use convolutional neural networks and autoencoders to denoise images. We will use the following points in this article to correctly understand this concept. 

Let's start by understanding noise.

Noise is usually defined as random changes in brightness or color information, and it is often caused by technical limitations of the image acquisition sensor or inappropriate environmental conditions. These difficulties are usually unavoidable in actual scenes, making image noise a common problem that must be solved by appropriate denoising methods.

Denoising an image is a difficult task, because noise is related to the high-frequency content (ie details) of the image. Therefore, the goal is to strike a balance between suppressing noise as much as possible and not losing too much information. The most commonly used filter-based image denoising methods, such as inverse filter, median filter and Wiener filter.

The presence of noise in the image may be additive or multiplicative. In the additive noise model, the additive noise signal is added to the original signal to produce a corrupted noise signal that follows the following rules:

w(x,y) = s(x,y) + n(x,y)

Here, s(x,y) represents the original image intensity, and n(x,y) represents the noise provided to produce a corrupted signal w(x,y) at the pixel location (x,y). Similarly, the multiplicative noise model multiplies the original signal by the noise signal.

In the process of image acquisition and transmission, noise may be introduced into the image. The introduction of noise into an image may be caused by many factors. The quantization of noise depends on the number of damaged pixels in the image. 

Image noise can range from almost invisible spots on digital snapshots taken under good lighting to optical and radio astronomy images that are almost entirely noisy, and a small amount of information can be extracted through complex processing. This level of noise is inappropriate in photos because the subject cannot be identified.

The following are the main sources of noise in digital images: – 

The pattern of noise and its probability characteristics can distinguish it. There are many types of noise. Although we mainly focus on the most important forms, they are Gaussian noise, salt and pepper noise, poison noise, impulse noise, and speckle noise.

As we all know, Gaussian noise is statistical noise whose probability density function (PDF) is equal to a normal distribution. Gaussian noise has a uniform distribution in the entire signal. 

The pixels of a noisy image are composed of the sum of their original pixel values ​​plus random Gaussian noise values. The probability distribution function of the Gaussian distribution has a bell shape. Additive white Gaussian noise is the most common application of Gaussian noise.

The following figure shows the Gaussian distribution function (probability distribution function) of Gaussian noise and the pixel representation of Gaussian noise. 

A common type of noise in photos is salt and pepper noise. It appears as white and black pixels appearing at random intervals. Errors in data transmission can cause this form of noise. The values ​​of a and b in salt and pepper noise are different. The average probability of each person is less than 0.1. The damaged pixels are alternately set to minimum and maximum values, giving the image a "salt and pepper" appearance. The distribution and pixel representation of this noise are shown below.

Using median filter, morphological filter or anti-harmonic averaging filter is an effective strategy to eliminate this type of noise. In the event of fast transients (such as improper switching), salt and pepper noise will spread into the image.

Unlike Gaussian noise or salt and pepper noise, speckle noise is multiplicative noise. In diagnostic examinations, this reduces image quality because it makes the image appear as backscattered waves caused by many microscopic scattered reflections flowing through internal organs. This makes it more difficult for the observer to distinguish the fine details in the image. The distribution and pixel representation of this noise are shown below.

This type of noise can be found in various systems, including synthetic aperture radar (SAR) images, ultrasound imaging, and more.

Poisson noise is generated by the nonlinear response of the image detector and recorder. This type of noise is determined by image data. This expression is used because the detection and recording procedure combines arbitrary electron emission with a Poisson distribution and average response value. Since the mean and variance of the Poisson distribution are the same, if the variance of the noise is assumed to be 1, the image correlation term is considered to have a standard deviation.

Deep learning-based technology has become the most successful solution to many real-world challenges that require digital image processing, and is also used as a natural alternative to non-learning correlation filters and denoising algorithms based on prior knowledge. This learning-based strategy is less affected by the non-linear characteristics of the noise generation process.

For a long time, Multilayer Perceptron (MLP) has been one of the most widely studied image denoising technologies based on machine learning. Due to the latest developments in computer graphics processing capabilities, especially image processing, MLP has been replaced by Convolutional Neural Networks (CNN).

Attention learning is a very interesting deep learning training method, and the image denoising has not been extensively studied. This technique can direct the learning of deep neural networks to more information components of the input data. The advantages of this method have achieved many breakthroughs in natural language processing, recommendation systems, healthcare analysis, audio recognition, and image classification.

Recently, the research paper Attention-Residual Convolutional Neural Network proposed a denoising technology based on deep learning, which combines the CNN model with the residual connection and attention mechanism.    

The above is the proposed architecture, where In is the input noise image, Id is the output denoised image, Conv and BN are the convolution and batch normalization layers respectively, and A1...A20 are the attention weights.

Once the attention residual mechanism (shown as a dashed rectangle) estimates the noise present in the image In, it can then be further eliminated from the image using a simple addition process to produce an Id denoised image, as shown in the above building. For more details about this architecture, please refer to the research. 

Another commonly used denoising method is to use an autoencoder, which is an artificial neural network that is mainly used to compress and decompress data by using encoders and decoders in a supervised manner. To use an autoencoder for denoising, train the encoder and decoder using the characteristics of the noisy image, and target a clean image. This form of method can give you fast and satisfactory results. I encourage you to read this article on autoencoders, where I discuss the usage of some applications.    

Through any practice or precise capture, there will be many images that need to be distilled so that we can extract as much information as possible. For this reason, we have seen the common types of noise and their importance. In addition, we also discussed how to use CNN to denoise images. 

Vijaysinh is a fan of machine learning and deep learning. He is good at machine learning algorithms, data manipulation, processing and visualization, and model building.

Copyright Analytics India Magazine Pvt Ltd