A Deep Learning-Based Circular End Face Detection and Sub-Pixel Centering Method

Abstract

This application provides a deep learning-based method for detecting circular end faces and its sub-pixel centering, comprising: acquiring images of circular end faces under different scenarios to form a training set; training the model using object detection and super-resolution to obtain a first network model; using the model to identify the training set to obtain a first image; acquiring the super-resolution image of the first image to obtain a second image; calculating a threshold for the second image and performing image binarization; setting connected components and calculating the weight w of each pixel in the second image. i Based on the weight w of each pixel i And the coordinates P of each corresponding pixel. i (x, y), calculate the center coordinates O0(x0, y0) of the second image; assume that the sub-pixel precision center point of each image in the training set is O c (x c ,y c ), calculate the sub-pixel precision center point O of each image in the training set based on the center coordinates O0(x0,y0). c (x c ,y c This application improves the image resolution and achieves sub-pixel level detection accuracy by using local super-resolution processing on the target circle.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a deep learning-based method for detecting circular end faces and its sub-pixel centering. Background Technology

[0002] Target detection and localization via camera is one of the key technologies in the field of deep learning. For a circular end face, the center of the circle is the localization feature point, so the accurate localization of the center is a crucial step.

[0003] Currently, the traditional methods for detecting the center of a circle mainly include the Hough transform method, the centroid method, and the least squares method. The Hough transform method is greatly affected by the degree of discretization of the parameter space. The centroid method requires a high degree of uniformity in the gray-level distribution of the image. The least squares method fits the edge points of the image, so the accuracy of edge point extraction directly affects the accuracy of the center of the circle.

[0004] Currently, the main methods for extracting subpixel edges include the moment method, interpolation method, and fitting method. Interpolation and fitting algorithms have high complexity and large computational load, resulting in long running time. The moment method has low computational accuracy, so its stability is not high. Summary of the Invention

[0005] The purpose of this application is to address the above problems by providing a deep learning-based method for circular end face detection and its sub-pixel centering.

[0006] This application provides a deep learning-based method for circular end face detection and its sub-pixel centering, including:

[0007] The camera was used to collect original images of the circular end face under different scenes to form a training set.

[0008] The training set is trained using object detection and super-resolution to obtain a first network model;

[0009] The first network model is used to identify images in the training set to obtain a first image;

[0010] Obtain the super-resolution image of the first image to obtain the second image;

[0011] Calculate the threshold of the second image and perform image binarization.

[0012] Set up connected components and calculate the weight w for each pixel in the second image. i ;

[0013] Based on the weight w of each pixel i And the coordinates P of each corresponding pixel. i(x,y), calculate the center coordinates O0(x0,y0) of the second image;

[0014] Assume that the sub-pixel precision center point of each image in the training set is O. c (x c ,y c The sub-pixel precision center point O of each image in the training set is calculated based on the center coordinates O0(x0,y0). c (x c ,y c ).

[0015] According to the technical solution provided in the embodiments of this application, the training set can be used for both target detection and super-resolution of the target region.

[0016] According to the technical solution provided in the embodiments of this application, the method used for target detection is the YOLO4 model.

[0017] According to the technical solution provided in the embodiments of this application, the method used for super-resolution is: EDSR model.

[0018] According to the technical solution provided in the embodiments of this application, the process of using the first network model to identify the training set includes: assuming that the side length of the camera's recognition box is m, cropping each image in the training set using the camera to obtain a cropped image of size m×m, ensuring that each image in the training set is in the cropped image, and that the cropped image does not include other circles, thus obtaining the first image, wherein the image coordinates of the first image at the upper left corner of the original image are L(x,y).

[0019] According to the technical solution provided in the embodiments of this application, the resolution of a local portion of the first image is magnified by n times to obtain the second image.

[0020] According to the technical solution provided in the embodiments of this application, the step of calculating the center coordinates of the second image includes:

[0021] The standard equation of a plane circle is (xa 2 )+(yb 2 ) = r 2 Set the coordinates P of each pixel. i Substituting (x,y) and O0(x0,y0) into the standard equation, we get:

[0022]

[0023] In the formula: i is a natural number greater than or equal to 1.

[0024] According to the technical solution provided in the embodiments of this application, the coordinates O of the center point of the circular sub-pixel precision can be obtained by the following formula based on O0(x0,y0). c (x c ,y c );

[0025]

[0026] In the formula: n represents the super-resolution magnification factor;

[0027] L x This represents the x-coordinate of the first image at the top left corner of the original image;

[0028] L y This represents the ordinate of the first image at the top left corner of the original image.

[0029] Compared with the prior art, the beneficial effects of this application are as follows: This application uses a camera to acquire images of a circular end face and form a training set. It then uses object detection and super-resolution methods to train the training set to obtain a first network model. The application also identifies images within the training set, acquires super-resolution images of the first image to form a second image, calculates its threshold and performs binarization, sets connected components, calculates the weight of each pixel in the second image, and then calculates the center coordinates of the second image based on the weights and coordinates of each pixel. Finally, it calculates the sub-pixel precision center point coordinates of each image in the training set based on the center coordinates. During use, the camera is first used to capture images of different fields... Images are acquired from the circular end face of the scene, and then used to form a training set. The threshold is calculated and binarized. Connected components are set and the weight of each pixel in the second image is calculated. Then, the center coordinates of the second image are calculated based on the coordinates of each pixel and its weight. Based on the center coordinates, the coordinates of the sub-pixel precision center point of each image can be obtained. The method used in this application achieves sub-pixel precision centering through deep learning. Compared with interpolation and fitting algorithms, its complexity and computational load are relatively small, resulting in a shorter running time. Compared with the moment method, the computational accuracy is improved and the stability is better. Attached Figure Description

[0030] Figure 1 This is a deep learning-based super-resolution model in the embodiments of this application;

[0031] Figure 2 This is a system flowchart of the method in the embodiments of this application. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solution of this application, the application will be described in detail below with reference to the accompanying drawings. The description in this section is only exemplary and explanatory, and should not be used to limit the scope of protection of this application.

[0033] Please refer to Figures 1-2 This embodiment provides a deep learning-based method for circular end face detection and its sub-pixel centering, including:

[0034] The camera was used to collect original images of the circular end face under different scenes to form a training set.

[0035] Specifically, the same camera is used to capture images of circular end faces in different scenes to obtain original images, and the training set is constructed based on the original images. The training set can be used for both object detection and super-resolution of the target region. That is, the training set can be used for both object detection and super-resolution at the same time. When using it, there is no need to construct different training sets for different targets. Only one training set needs to be constructed initially to complete both object detection and super-resolution.

[0036] The training set is trained using object detection and super-resolution to obtain a first network model;

[0037] Specifically, the object detection method is the YOLO4 model; the super-resolution method is the EDSR model. The YOLO4 model and the EDSR model are used to train each image in the training set to obtain a first network model. The first network model is the network model formed after training each image with the YOLO4 model and the EDSR model. Both the YOLO4 model and the EDSR model belong to existing technologies.

[0038] The first network model is used to identify images in the training set to obtain a first image;

[0039] Specifically, the first network model will recognize each image in the training set separately. During the recognition process, assuming the side length of the camera's recognition box is m, the camera will crop each image in the training set to obtain a cropped image of size m×m. During the recognition and cropping process, it must be ensured that each image in the training set is cropped. If there is only one image in the training set, it must be ensured that this image is within the camera's recognition box. If there are three images in the training set, it must be ensured that the camera's recognition box can recognize and crop each of the three images separately. Each cropped image contains only one image (i.e., after cropping, each image recognized by the recognition box can only contain one image, i.e., only one circle), and cannot contain a second or third image (only one circle can appear in the image during the recognition process and after cropping, and two or three circles cannot appear simultaneously). After recognizing each image, a first image is obtained, and the image coordinates of the first image at the upper left corner of the original image are denoted as L(x,y).

[0040] Obtain the super-resolution image of the first image to obtain the second image;

[0041] The resolution of a portion of the first image is magnified by n times to obtain the second image.

[0042] Set up connected components and calculate the weight w for each pixel in the second image. i ;

[0043] After obtaining the second image, connected components are set for the second image, and the weight w of each pixel in the second image is calculated. i And calculate the weight w of each pixel in the second image. i .

[0044] Based on the weight w of each pixel i And the coordinates P of each corresponding pixel. i (x,y), calculate the center coordinates O0(x0,y0) of the second image;

[0045] Each pixel has its own coordinates P i (x,y) and weight w i Based on the coordinates P of each pixel i (x,y) and weight w i Calculate the center coordinates O0(x0,y0) of the second image, and the weight w i The calculation method is existing technology and will not be elaborated here.

[0046] The standard equation of a plane circle is given by (xa) 2)+(yb 2 ) = r 2 Set the coordinates P of each pixel. i Substituting (x,y) and O0(x0,y0) into the standard equation, we get:

[0047]

[0048] In the formula: i is a natural number greater than or equal to 1.

[0049] Assume that the sub-pixel precision center point of each image in the training set is O. c (x c ,y c The sub-pixel precision center point O of each image in the training set is calculated based on the center coordinates O0(x0,y0). c (x c ,y c );

[0050] The formula for calculating the sub-pixel precision center point is expressed as follows:

[0051]

[0052] In the formula: n represents the super-resolution magnification factor;

[0053] L x This represents the x-coordinate of the first image at the top left corner of the original image;

[0054] L y This represents the ordinate of the first image at the top left corner of the original image.

[0055] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A method for detecting circular end faces and its sub-pixel centering based on deep learning, characterized in that, include: The camera was used to collect original images of the circular end face under different scenes to form a training set. The training set is trained using object detection and super-resolution to obtain a first network model; The first network model is used to identify images in the training set to obtain a first image; Obtain the super-resolution image of the first image to obtain the second image; Calculate the threshold of the second image and perform image binarization. Define the connected components and calculate the weight of each pixel in the second image. wi ; Based on the weight of each pixel wi and the coordinates of each corresponding pixel. Pi ( x , y ), calculate the center coordinates of the circle in the second image. O 0( x 0, y 0); Assume that the sub-pixel precision center point of each image in the training set is Oc ( xc , yc According to the coordinates of the center of the circle O 0( x 0, y 0) Calculate the sub-pixel precision center point of each image in the training set. Oc ( xc , yc ); The process of using the first network model to recognize the training set includes: assuming the side length of the camera's recognition box is m, cropping each image in the training set using the camera to obtain a cropped image of size m×m, ensuring that each image in the training set is within the cropped image, and that the cropped image does not contain other circles, thus obtaining the first image. The image coordinates of the top-left corner of the first image in the original image are... L ( x , y ); The formula for calculating the sub-pixel precision center point is expressed as follows: In the formula: n represents the super-resolution magnification factor; Lx represents the x-coordinate of the first image at the top left corner of the original image; Ly represents the vertical coordinate of the first image at the top left corner of the original image.

2. The method for detecting circular end faces based on deep learning and its sub-pixel centering according to claim 1, characterized in that, The training set can be used for both object detection and super-resolution of the target region.

3. The deep learning-based circular end face detection and sub-pixel centering method according to claim 1, characterized in that, The method used for target detection is the YOLO4 model.

4. The method for detecting circular end faces based on deep learning and its sub-pixel centering according to claim 1, characterized in that, The method used for super-resolution is the EDSR model.

5. The method for detecting circular end faces based on deep learning and its sub-pixel centering according to claim 1, characterized in that, The resolution of a portion of the first image is magnified by n times to obtain the second image.

Images