What Is Microscopic Image Processing?

A microscopic image is an image observed in a microscope. With the development of computer image processing technology and pattern recognition, the analysis and processing of microscopic images has gradually been applied in scientific research. One of the most important aspects is the classification and identification of microorganisms. ^[1]

The geometric features of a single cell are the geometric features of each connected area, and there is a close relationship and influence between each feature quantity, and the distribution and statistics of each cell in the sample is the distribution and statistics of each connected area in the image. Calculation is the basis of statistics and analysis, and correct judgment and statistics need to be obtained through further calculation and analysis of mathematical statistics, mean clustering, etc. The calculation of each connected domain includes the calculation of geometric features such as perimeter, area, and center of gravity. Morphological features such as roundness, rectangleness, application of moments, calculation of edge complexity. ^[2]

Medical microscopic images are an important reference for medical diagnosis, and also an important part of modernization and computerization of medical methods. The analysis of traditional medical microscopic images is generally performed by the patient under the light microscope after the sample is sent for examination The inspector views and counts and analyzes with the naked eye, and gives a diagnosis conclusion against the sample micrograph. Such an operation process makes the inspector's work more intensive, requires higher technical level of the inspector, and the test result is Recording in writing is not conducive to the further analysis of the microscopic images by the attending physician and the inpatient physician.

If computer technology is used to ingest and record microscopic images in the form of digital images, the use of this medical information will be greatly improved. As the inspectors observe the microscopic images on a large screen, the work intensity will be greatly reduced. The microscopic images will be printed on the diagnostic report, providing an image sample for further analysis to other physicians and patient families.This is also a shared resource on the hospital medical information network, which is also part of the Internet-based long-distance medical diagnostic system. As the standard micrograph sample has been stored on the computer, the inspector can make analysis and discrimination against the measured legend on the computer. At the same time, it is added to the software system.

The main problem of image stitching is to deal with the overlap between two adjacent image samples. The quality of image stitching mainly depends on the degree of alignment of the images. The usual matching algorithm is based on the gray similarity of the corresponding pixels of the overlapping parts of the two images in the RGB or CMY color system.

A fast matching algorithm based on grid matching has been proposed.The basic idea is to take advantage of the small difference in the gray values of adjacent pixels in the image, take a grid in the overlapping area of the second image, and then The grid is moved on the first image, and the sum of the squares of the differences between the corresponding RGB values of the two images of all grid points is calculated. The grid position where the smallest value is recorded is considered as the best matching position. And the matching is divided into two steps of thickness, which can significantly reduce the amount of calculation.

Image stitching technology is currently mainly used in panoramic video systems and geographic information systems to realize the stitching of several partially overlapping small images into one large image to obtain panoramic information. In the study of microscopic images, there are similar problems in image stitching. Intelligent ferrography diagnostic system is a typical application field of microscopic images ^[4] . The basic idea is to identify mechanical wear particles in microscopic images for mechanical wear failure analysis. At low objective magnifications (such as × 4, × 10), the overall shape of the abrasive particles can be displayed in the same scene, but at high magnifications (such as × 40, × 60), A larger size abrasive particle is often not displayed in the field of view, so the phenomenon of truncated abrasive particles often occurs, which leads to difficulties in subsequent identification of abrasive particles. ^[5]

Micrograph template matching stitching

Template matching, also known as block-kuai matching, is one of the most commonly used image matching algorithms. Make both matched expressions images in template matching. The matching of the image can be the matching between the entire image and the entire image, or a part between the images (sub-images) .Obviously, the template matching in image stitching is to establish a template in the second image. In the first image, find the image area that is most similar to the template, so as to determine the overlapping area of the image, and finally achieve image stitching. The main work in matching is to translate the template on the graph and calculate the correlation. The place with the largest correlation value is the best match. ^[5]

Template matching diagram

Micrograph multi-resolution matching

Also known as multi-resolution matching based on pyramid data structure. In template matching, if the size of the template is selected too large, sufficient accuracy can be guaranteed, but the speed of matching will be greatly reduced; if the size of the template is selected too small, although the calculation speed can be effectively improved, the accuracy of the matching will be reduced. Obvious reduction.

An effective method to reduce the matching search time is to use the pyramidal data structure of the image to find the matching position in the order of first coarse and then fine.

Pyramid image data structure

The basic idea is to first average the gray values of the pixels in the 2 × 2 domain in the image to obtain an image with a lower resolution. According to this principle, a set of pyramid-shaped images can be obtained. After preprocessing both the two pictures to be matched, the matching search starts from a certain low resolution of the two pictures. Not only the number of image points is small, but the high-frequency information is also smoothed out. Therefore, a rough match result may have more than one match position. Because there are few points in the figure, even if the rough matching is performed on all reference points, the search process is still fast. The next step is to find a match on the map with a higher level of resolution, but the search space is limited to one or a few coarse matching points, and the calculation is not large. The third match is the same as the second. This continues until the matching point of the second image is found on the original image of k = 0.

When applying the multi-resolution matching technology based on the image pyramid data structure to image stitching, note the following:

(1) The number of layers of the pyramid data structure of the image is not as good as possible, it must have an optimal value. Generally, the topmost image data has a length and width of not less than 200 pixels.

(2) Because the top level of the image's pyramid data structure has low resolution and low matching accuracy, the best matching point on the top level is not necessarily the true best point. If you only find a best matching point on the top level, Then the search is performed on the layer-by-layer basis based on the top-level matching point, so the probability of mismatch will be greatly increased. The solution is to use the size of the correlation coefficient at the top level to find multiple matching points from large to small, and then search down layer by layer based on the multiple coarse matching points, so the reliability of the matching will be greatly improved Therefore, the number of rough matching points needs to be carefully considered. Generally take 5 to 10 points.

(3) Since the image matching algorithm based on block matching is highly dependent on the image gray level, the difficulty of image matching of different qualities is different. For images with clear images, strong contrast, and many scenes in the image, it is obviously easy to match. At this time, it is required that the image overlap area is smaller and the template can be smaller; for images with blurred and poor contrast, especially for images with few scenes in the image, stitching will become difficult and sometimes even fail. The image overlap should be more, and the matching template should be larger. Therefore, when stitching specific images, the size of the overlapping areas of the images and the size of the matching template need to be carefully considered. Generally, the overlapping area of the image is required to be about 1/2 images wide. ^[5]

Seamless stitching of micrographs

Image stitching diagram

When the best matching point of the two images to be stitched is found, the next step is to combine the two images into one image. If the overlapping part is simply to take the first or second image data in the overlapping area, It will cause image blur and obvious connection traces, which cannot be tolerated by image stitching. It will also affect subsequent image segmentation and recognition. Even if the average of the overlapping areas of two images is taken, the effect is not satisfactory. Therefore, in order to make the stitching area smooth, a fading method is widely used at present, that is, the first image is slowly transitioned to the second image in the overlapping portion. This allows seamless stitching of images. The left-right stitching of the two images is shown in the figure. ^[5]