What Is the Recovery Model?

Image restoration is to use the prior knowledge of the degradation process to restore the original appearance of the degraded image.

Image restoration technology is mainly proposed for the "degeneration" in the imaging process, and the "degradation" phenomenon in the imaging process mainly refers to the imaging system is affected by various factors, such as the defocus of the imaging system, the existence of equipment and objects The relative motion or the inherent defects of the equipment, etc., cause the image quality to fail to meet the ideal requirements. There are similarities between image restoration and image enhancement, and it is also to improve the overall quality of the image. However, compared with image restoration technology, image enhancement technology focuses on the stretching of contrast. Its main purpose is to process the image according to the viewer's preferences and provide the image to the viewer. The image restoration technology is The deblurring function removes the blurred parts in the image and restores the authenticity of the image. The main method is to use some so-called a priori knowledge of the degraded image to repair or reconstruct the degraded image. From the perspective of the restoration process, it can be regarded as a reverse process of image degradation. For the restoration of an image, an appropriate estimation of the entire process of image degradation must first be made, and an approximate mathematical model of degradation should be established on the basis of this. After that, the model needs to be appropriately modified to compensate for the distortion in the degradation process to ensure The image obtained after the restoration approaches the original image, and optimizes the image. But in the process of image degradation and blurring, noise and interference coexist, which brings a lot of uncertainty to the image restoration.

From the above overview of image restoration technology, we can see that the basis of image restoration technology should be a mathematical model of image degradation, and the image degradation models that different imaging systems have are different.

(1) Classification of image noise. We usually divide the noise that affects image quality into four basic types: one is that those recorded on the photosensitive film are susceptible to the noise fluctuations of the photosensitive particles; the other is that when the image is converted from optical to electronic form In the process, it is performed in a statistical form. This is mainly because the number of photons received by each pixel is random and limited, which leads to the generation of optoelectronic noise. Third, the electrons The amplifier also introduces thermal noise in the process of processing the signal. Fourth, in the process of acquiring images, it is easy to obtain periodic noise from electricity or electromechanical interference.

(2) Establish the corresponding probability density function according to the characteristics of the image. In the process of processing digital images, it is generally necessary to use the probability density function as the fundamental basis to express the statistical characteristics of noise and establish a corresponding data model. There are six common types of typical noise: Gaussian noise; sharp noise; gamma noise; exponential distribution noise; uniform distribution noise and impulse noise. Due to space limitations, the probability density functions of various noises are not introduced here.

(3) Estimating relevant parameters of noise. In the process of noise processing, various noise related parameters need to be obtained, which is part of grasping the prior index of the image. For example, Wiener filtering (ie, least mean square) method is used to restore the image, and Kalman filtering is used to smooth the image, edge detection, and segmentation of the image, etc. All need to estimate the variance of noise. Under normal circumstances, since the main data is the degraded image, this is only a blind estimation of the noise variance. The estimation of noise variance mainly includes the following two types: first, pre-filtering the noisy image, and then performing variance estimation on the filtered data; second, dividing the noise into multiple regions before the estimation, The noise variance is estimated mainly for the "flat area". The commonly used estimation methods are: average method, median method, block method, scatter method, pyramid method and pre-filter method.

(4) Remove noise. A basic theory of noise removal is to propose a low-pass filtering method based on the high-frequency characteristics of noise. Common removal methods include mean filtering and median filtering. Although these two methods can remove noise, they also remove relevant details of the image, resulting in blurred image boundaries. Therefore, a model-based denoising algorithm is now proposed, which is mainly based on the Markov model of the image itself and different noises.