图像去噪声

4.1.5 Image noise removal
4.1.5.1 Common image noise
Another step in the pre-processing is the elimination of noise. First of all, introduce some common noise:

1. The Gaussian noise, that is the number of nodes on a certain block or blocks, has the largest amount of noise in the graph, and centering on this point, the distribution is spread around, and the noise is getting less and less. We generally use the mean algorithm to eliminate this kind of noise, the specific algorithm will be introduced below. The reason is the market for the image sensor is not bright enough and the brightness is not uniform enough when shooting, the components of the circuit have their own noise and mutual influence and the image sensor has been working for a long time, and the temperature is too high.
2. Salt and pepper noise is black and white light and dark point noise generated by image sensors, transmission channels, and decoding processes. Salt and pepper noise refers to two kinds of noise, one is salt noise and the other is pepper noise [21].
3. Multiplicative noise exists in real-world image applications, such as coherent image systems such as synthetic aperture radar, ultrasound, and laser [22]. The multiplicative noise notifies the direct impact of the received signal. Without the signal, the multiplicative noise would not exist.

4.1.5.2 Filtering and Denoising
First, let us describes convolution. Template convolution is a neighbourhood computing method commonly used in digital image processing. Template convolution can implement functions such as image smoothing, image sharpening, and edge detection. The template can be a small image or a filter. In template convolution, When we get the template, the image is divided into pixel regions with the same size as the template, and then the content of the template and the content in the pixel region are compared. The result of the comparison generates a percentage of similarity. Finally, the result is given to the input image, whose pixel position is consistent with the position of the template centre on the input image.

The concept of median filtering is very simple: the grey levels of all pixels in a small window centred on the current pixel are sorted from small to large, and the middle value of the sorted result is used as the grey value of the pixel.

The processing idea of the mean filter is also very simple. We need calculating the offset of the baseline in the window area, and then replacing the corresponding calculation to the original intermediate point. This algorithm has the advantages of high efficiency and simple thinking. Similarly, the disadvantages are also obvious. Calculating the mean value will “blur” the edge information and feature information in the image, and many features will be lost. Moreover, there are many optimization methods that can be used to calculate the mean filtering. For example, the integral image is used to preprocess the image. The processed image can obtain the sum of pixels in the window area by O (1) time complexity. If we use parallel and SSE instruction set for acceleration, the effect will be faster.

4.1.5.3 Non-local Mean

Non-local mean filtering is a filtering method proposed by Buades et al. In 2005 [23]. The basic idea is to calculate the weight of neighborhood pixels based on the self-similarity of the image [23]. The method of smoothly averaging the area around a target pixel, The gray value of the current pixel is obtained by weighting the average value of all the pixels with similar structure in the image. In order to measure the pixels with similar structure, the algorithm uses the image sub-block (7 * 7) or (9 * 9) with the current pixel as the center for each pixel’s weight Gaussian weighted Euclidean distance betwee n them.The filtered image has high definition without losing detail. When Baudes put forward, the starting point should be to borrow the phenomenon that the more images are weighted, the better the effect. Then, in the same image, classify regions with the same properties and weight the average to obtain a denoised image. Noise reduction should be performed. The effect will be better. Compared with bilateral filtering, non-local mean filtering determines the similarity of two pixels based on the similarity between image pieces of a certain size. Compared with bilateral filtering, it can better identify the details of the image. However, the filtering parameter h of the non-local mean algorithm is difficult to adjust. H controls the speed of attenuation. If there is no too large attenuation algorithm, it will degenerate to mean filtering. If the attenuation algorithm is too small, there will be no smoothing effect. Generally, the selection of h parameter will refer to Standard deviation of image noise. The filtering process of NL-Means can be expressed by the following formula:

[38]
w stands for weight. There are many methods to measure similarity, the most commonly used is to estimate based on the square of the brightness difference between two pixels. Due to the noise, a single pixel is not reliable, so using their neighborhood, only when the neighborhood has a high degree of similarity can it be said that the two pixels have a high degree of similarity.