Digital Image Processing

Title: Target detection and classification in large-scale images

In the image target recognition and tracking, the image captured by the camera, in the imaging, digitization and transmission process, will inevitably be affected by a variety of noise interference, the quality of the image will often appear unsatisfactory degradation, affecting the visual effect of the image. Usually these noise interference causes the image degradation, displays the image blurred, the characteristic drowns, this may cause the image analysis to be unfavorable, causes the obtained image quality to be low. It is more difficult to identify and track the target directly in such images. It is very important to suppress the various interfering signals that degrade the image, to enhance the useful signals in the image, and to calibrate the observed images under the same constraint condition.

The purpose of the filter is two: one is to extract the characteristics of the object as a feature pattern of image recognition, and the other is to adapt to the requirements of image processing to eliminate the noise mixed in the image digitization.

Gaussian filtering is a linear smoothing filter, which is suitable for filtering Gaussian white noise, and has been widely used in the preprocessing stage of image processing. According to my understanding, the image is Gaussian filtering is the image of each point of the pixel value calculation, the calculation criterion is the gray value of the point itself and other pixels in its neighborhood weighted average, and weighted average weight coefficient by the two-dimensional discrete Gaussian function sampled and normalized after the resulting.

Mean filter is also called linear filter, and its main method is neighborhood averaging method. The basic principle of linear filtering is to replace each pixel value in the original image with the mean, that is, to treat the current pixel (x, y) of the processing, select a template that consists of several pixels of its nearest neighbor, the mean value of all the pixels in the template, and then assign the mean to the current pixel (x, y). As the grayscale value g (x, y) of the image at that point after processing, g (x, Y) =1/m∑f (x, y), M is the total number of pixels in the template that contain the current pixel. This method can smooth the image, the speed is fast, the algorithm is simple. But the noise can not be removed, which weakens it weakly.

median filtering is a nonlinear smoothing technique that sets the grayscale value of each pixel point to the median of all pixel grayscale values in a neighborhood window of that point. The implementation process is:1) Sort by extracting odd numbers of data from a sample window in the image2) Using the sorted median as the gray value of the current pixelIn image processing, median filtering is often used to protect edge information, which is a classical method of smoothing noise, which is very effective to eliminate salt and pepper noise, and has special effect in the phase analysis and processing method of optical measurement fringe image, but it has little effect in the Fringe Center analysis method.

Bilateral filtering is a filter that can be used for edge-preserving denoising. This effect can be achieved because the filter is composed of two functions, one function is determined by the geometric space distance filter coefficients, the other one by the pixel difference determines the filter coefficients.

In the previous several methods of filtering, the gray value of pixel point is determined by the gray value of other points in the neighborhood of the point, such as Gaussian filter and mean filter can be regarded as weighted average, median filter is the neighborhood gray value. The bilateral filter not only takes into account the gray value of the point in the neighborhood, but also considers the geometric distance of the point distance from the center point, so that the gray value expression formula of the filtered point can be obtained:

where k is the normalized coefficient, the expression is:

H and X are the gray values of the corresponding points after filtering and filtering respectively

c indicates the spatial similarity between the center point and the point within its neighborhood

s represents the grayscale similarity of points within the center point and its neighbors.

In the implementation process, the C and S functions are implemented using the Gaussian function, which is defined as follows:

In terms of enhancement, the most important concern is the average and variance of an image's grayscale (a measure of the average contrast).

The method of judging whether a point is dark or bright is to compare the local average grayscale with the average grayscale of the image.

The concept of the filter is derived from the Fourier transform which is processed in the frequency domain comparison signal.

The filtering of nonlinear space is based on domain processing.

A smoothing spatial filter is used to obfuscate and reduce noise.

Sharpening the spatial filter, the Laplace operator.

Digital Image Processing