With the increasing use of digital cameras, people come across a wide variety of images in their daily life. Some of the images are of good quality while some of the images are of poor quality. The image quality also deteriorates due to the presence of noise. This noise can be caused by poor lighting or other intensity issues. There are various approaches to denoise an image, ie to reduce the noise in an image. It’s been a hot research topic for a long time and is still being experimented with by researchers. Here we will discuss how convolutional neural networks and autoencoders are used to denoise an image. We’ll go through the following points in this article to properly understand this concept.
table of contents
- What is noise
- Sources of noise
- Different types of sounds
- Noise reduction with CNN
- Image noise with auto encoders
Let’s start by understanding the sound.
What is noise
Noise is typically defined as a random variation in the brightness or color information and is often generated by technical limitations of the image capture sensor or by unsuitable environmental conditions. These difficulties are often inevitable in real-life scenarios, making image noise a common problem that needs to be addressed with appropriate noise reduction approaches.
Dismissing an image is a difficult task because the noise is tied to the high-frequency content of the image, ie to the details. Therefore, the aim is to find a balance between the greatest possible suppression of noise and the simultaneous loss of information. Filter-based approaches to image noise reduction, such as the inverse, median, and Wiener filters, are the most commonly used.
The presence of noise in an image can be additive or multiplicative. In the additive noise model, an additive noise signal is added to the original signal in order to generate a corrupted noise signal that follows the following rule:
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 that is added to generate the corrupted signal w (x, y) at (x, y) pixel position. Similarly, the multiplicative noise model multiplies the original signal by the noise signal.
Sources of noise
Noise can be introduced into the image during image acquisition and transmission. Several factors can cause noise to be introduced into the image. The quantification of the noise is determined by the number of damaged pixels in the image.
Image noise can range from nearly invisible spots on a digital snapshot taken in good lighting to optical and radio astronomical images made almost entirely of noise from which, through complex processing, a small amount of information can be extracted. Such a level of noise would be inappropriate for a photo as it would be impossible to identify the subject.
The following are the main sources of noise in digital images: –
- Environmental factors can affect the image sensor.
- Low light conditions and sensor temperature can lead to picture noise.
- Dust particles in the scanner can cause noise in the digital image.
- Disturbance of the transmission channel.
Different types of sounds
The pattern of noise, as well as its probabilistic properties, distinguish it. There are a variety of types of sounds. While we mainly focus on the most important forms, these are Gaussian noise, salt and pepper noise, poison noise, impulse noise, and speckle noise.
It is well known that Gaussian noise is statistical noise with a probability density function (PDF) equal to the normal distribution. Gaussian noise has an even distribution over the signal.
A noisy image has pixels that are composed of the sum of their original pixel values plus a random Gaussian noise value. The probability distribution function for a Gaussian distribution has a bell shape. Additive white Gaussian noise is the most common use for Gaussian noise in applications.
The following figure shows the Gaussian distribution function (probability distribution function) of Gaussian noise and the pixel representation of Gaussian noise.
Salt and pepper sounds
One type of sound that is commonly seen in photos is the salt and pepper sound. It manifests as white and black pixels that appear at random intervals. Data transmission errors cause this type of noise. The values a and b in the salt pepper noise are different. Everyone has a probability of less than 0.1 on average. The damaged pixels are alternately set to the lowest and highest value, giving the image a “salt and pepper” appearance. The distribution and pixel representation of this noise is shown below.
Using a median filter, a morphological filter, or a contra harmonic mean filter is an effective noise suppression strategy for this type of noise. In situations where rapid transients, such as B. incorrect switching occur, salt and pepper noise creeps into the images.
In contrast to Gaussian or salt and pepper noise, speckle noise is multiplicative noise. In diagnostic examinations, this degrades the image quality by giving the images a backscattered wave image caused by many microscopically dispersed reflections flowing through internal organs. This makes it difficult for the viewer to distinguish fine details in the images. The distribution and pixel representation of this noise is shown below.
This type of noise can be found in a wide variety of systems including synthetic aperture radar (SAR) images, ultrasound images, and many more.
Poisson noise is generated by the non-linear responses from image detectors and recorders. This type of noise is determined by the image data. Since detection and recording methods involve any electron emission with a Poisson distribution and a mean response value, this term is used. Since the mean and the variance of a Poisson distribution are the same, assuming that the noise has a variance of one, the image-dependent term is assumed to have a standard deviation.
Noise reduction with CNN
Deep learning-based techniques have proven to be the most successful solutions to many real-world challenges that require digital imaging, and have also been used as natural surrogates for non-learning-dependent filters and well-known knowledge-based noise reduction algorithms. Such learning-based strategies are less influenced by the non-linear properties of noise generation processes.
Multilayer Perceptrons (MLPs) have long been one of the best-studied machine learning-based techniques for denoising images. MLPs have been replaced by Convolutional Neural Networks (CNNs) due to recent developments in computer graphics processing capability, particularly for image processing tasks.
Attention learning is a very interesting deep learning training approach that has not yet been fully explored for image noise reduction. Such a technique can direct the deep neural network learning efforts towards more informative components of the input data. The benefits of such a technique result in multiple breakthroughs in natural language processing, recommendation systems, health analysis, audio recognition, and image classification.
Recently the research paper Attention Rest Convolutional Neural Network proposes a deep learning-based noise suppression technique that incorporates the CNN model with residual connection and attention mechanisms.
Above is the proposed architecture where In is the noisy input image and I.D. is the denoise output image, Conv and BN are folding or stack normalization layers and A1 … A20 are the attention weights.
As soon as the attention rest mechanism (shown in a dashed rectangle) has estimated the noise present in the imagen, then it can be further removed from the image using a simple additive process, resulting in the ID. de-noised image as shown in the architecture above. For more details on this architecture, go through the research.
Image noise with auto encoders
Another commonly used approach to noise reduction is with autoencoders, an artificial neural network primarily used to compress and decompress data by using encoders and decoders in a supervised manner. To use auto encoders for noise reduction, train the encoders and decoders with noisy images for functions and cleaned images as targets. This type of approach can give you quick and satisfactory results. I encourage you to read through this autoencoder article where I discussed how to use it by covering some applications.
With every exercise or with precise recording, many images will have to go through the distillation process so that we can extract as much information as possible. For this we have seen frequently observed types of noise and their meaning. We also discussed how CNN can be used to remove noise from the picture.
Join our Discord server. Become part of a dedicated online community. Join here.
Subscribe to our newsletter
Get the latest updates and relevant offers by sharing your email.