Image sub-pixel edge extraction method having extensive adaptability

Abstract

The invention discloses an image sub-pixel edge extraction method having extensive adaptability, comprising steps of adopting an adaptive high-low threshold value calculation method, executing a local maximum central value choosing operation on a gradient image by combining with gradient direction information of a pixel point after obtaining a gradient image, establishing a relative coordinate with a random pixel point as an origin point, taking eight-neighborhood pixels around the origin point as local maximum central value choosing data samples, obtaining a comparison result between the adjacent neighborhoods according to a gradient direction, determining whether the current pixel position is a boundary point candidate position, adopting a Hessian matrix method based on a Steger surface fitting method to solve a sub-pixel position of the boundary point and connecting boundary points for forming a line to constitute a set of directed continuity points. Determination whether an extreme value of a local gradient magnitude is a boundary point can be made by combining with a specific threshold value; if the local extreme value is greater than a given threshold value, the local extreme value is marked as a boundary point, and if the local extreme value is smaller than a given threshold value, the local extreme value is a noise point or a background point The invention has great instantaneity.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to an image sub-pixel edge extraction method. Background technique [0002] In machine vision, in order to perform target positioning, measurement, detection or geometric feature extraction, etc., it is necessary to extract the edge of the target with sub-pixel accuracy. For example, the template matching method using geometric features in target positioning requires sub-pixel edge extraction for templates and targets; in measurement applications, it is necessary to accurately detect the edges of objects to perform accurate measurements; in detection applications, such as optical characters Verification of OCV, edge defect detection, etc. all require stable detection of sub-pixel edges of objects. [0003] Commonly used edge extraction algorithms include Roberts operator, Sobel operator, Prewitt operator, Laplacian operator and Canny operator, etc. Edge extraction algori...

AI Technical Summary

Problems solved by technology

The use of the Hessian matrix method is the same as the present invention, but there is an essential difference in the pixel precision position calculation, which also leads to completely different applicability of the two methods

[0005] However, in industrial environment applications, images are disturbed by various factors, resulting in reduced image quality, including strong noise, edge blur, etc. How to stably detect sub-pixel-accurate edge features in low-quality images has not been well resolved.

Traditional pixel-accurate edge extraction algorithms cannot meet the accuracy requirements in industrial automation applications such as 3C automation equipment, electronic manufacturing, and industrial robot vision.

Sub-pixel edge extraction algorithms such as spatial moment method, gray moment method, Zernike moment method and digital correlation method have their own shortcomings in detection accuracy, calculation speed and anti-noise ability, and it is difficult to adapt to the harsh detection work in industrial environments. condition

[0006] Patent Document 1 can only extract the sub-pixel position of an elliptical object, which is insufficient in versatility, and cannot deal with the edge extraction of blurred objects

In the method disclosed in Patent Document 3, Sobel, Canny and LoG operators are used for edge detection, and then the results of the three operators are used for weighted voting statistics, and the sub-pixel coordinates are obtained according to the voting weight matrix. The problems of this method It is slow, the accuracy depends on the weight matrix, and it cannot solve the edge extraction problems of images such as strong noise and blur

The methods disclosed in Patent Document 4 and Patent Document 5 use the results of Canny and relative variation, and use machine learning methods for edge extraction, which are slow and cannot perform stable edge extraction in low-quality images

The method disclosed in Patent Document 6 performs sub-pixel edge detection in 8 gradient directions on the basis of coarse positioning of pixel coordinates. This method has a good calculation speed, but does not consider the processing of strong noise and blurred images

Patent Document 7 and Patent Document 8 also have low computational efficiency and cannot handle the edge extraction of strong noise and blurred images

The method disclosed in Patent Document 9 adopts multi-level Gaussian convolution operation, and the algorithm complexity is high. It cannot realize the robust extraction of the centerline of the light bar for the non-linear shadow changes in different areas caused by illumination changes and other reasons. This method is only applicable to laser light. Extraction of the centerline of the strip, unable to achieve general image edge feature extraction

Images