CN-122023905-A - Deep learning classification method for visible light image surface defects

CN122023905ACN 122023905 ACN122023905 ACN 122023905ACN-122023905-A

Abstract

The invention relates to the technical field of image recognition, in particular to a method for deep learning and classifying surface defects of visible light images, which is used for collecting visible light image channel information of wind power blades, fusing brightness, extracting direction mutation position points, the extending direction track forms a continuous channel, the edge brightness change is extracted to generate a jump boundary area, the color direction is aligned to replace image content, and the continuous area is marked to output a defect classification result. According to the invention, the local difference of the brightness directions of pixels is extracted, the identification basis of abrupt points is enhanced, the structural change area is highlighted, the connecting fracture path is linearly extended by combining the direction track, the structural continuity is maintained, the brightness change trend is integrated to form a continuous boundary area, the area division contrast is enhanced, the channel color difference interference is reduced by unifying the color directions, the consistency of the color relationship is maintained, the cooperative support is formed by the structural edge color information, and the stability of the classification basis is improved.

Inventors

ZHANG HAIFENG
KONG WEIJIE

Assignees

新沂市合沟众鑫风力发电有限公司

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (8)

1. The deep learning classification method for the surface defects of the visible light image is characterized by comprising the following steps of: S1, collecting red, green and blue channel information in a wind power blade image, carrying out brightness fusion on channel content, processing brightness directions between each pixel and adjacent pixels in the image, extracting position points in a direction mutation area, and keeping pixel distribution in the image to obtain a direction change breakpoint position diagram; S2, connecting the direction tracks in the image areas extracted from the direction change breakpoint position diagram, extending the direction information of adjacent pixels to enable the direction interval areas to be continuous, and taking the processed image as a subsequent input to obtain direction path recombined image content; S3, referring to image areas in the direction path reorganization image content, extracting edge brightness expression, observing brightness changes among different areas, separating out image parts with obvious brightness change trend, processing the change trend concentrated area into continuous image edges, and outputting a processing image layer to obtain a brightness jump boundary area diagram; And S4, reading the image part extracted from the brightness jump boundary region image, respectively processing the color directions of the red channel, the green channel and the blue channel, keeping the color relation of the region consistent by adjusting the color directions, replacing the processed image part with the corresponding region of the original image, and outputting a color value correction image frame.
2. The method for deep learning and classifying defects on a visible light image surface according to claim 1, wherein the direction change breakpoint position map comprises breakpoint position coordinates, a breakpoint pixel point set and a breakpoint distribution density, the direction path reorganization image content comprises reorganization direction channels, a direction continuity constraint domain and a fracture completion region, the brightness jump boundary region map comprises jump boundary outlines, brightness difference values at two sides of the boundary and jump intensity classification, and the color value correction image frame comprises a channel color consistency relation, a color correction parameter set and corrected color distribution.
3. The method for deep learning and classifying visible light image surface defects according to claim 1, wherein the step of obtaining the direction change breakpoint position map comprises the following steps: s101, collecting information of a red channel, a green channel and a blue channel in a visible light image of a wind power blade, respectively extracting brightness data of each channel at a pixel position, calling brightness information of corresponding positions of three channels, and fusing to generate a comprehensive brightness value of each pixel of the image to obtain a fused brightness matrix; s102, extracting the brightness difference between each pixel and the adjacent pixels according to the pixel brightness values in the fusion brightness matrix, identifying pixel points with brightness change exceeding a threshold value in each direction, screening out position points with direction mutation characteristics, and obtaining a direction change position set; And S103, extracting corresponding pixel information from the original image according to the image coordinates in the direction change position set, constructing a pixel distribution diagram only retaining the direction mutation position, and generating image space distribution of the marked mutation region to obtain a direction change breakpoint position diagram.
4. The method for deep learning and classifying visible light image surface defects according to claim 1, wherein the step of obtaining the direction path reorganized image content is as follows: S201, extracting the starting position and the ending position of a direction track based on the pixel region extracted from the direction change breakpoint position diagram, calling the spatial distribution of each pixel in the image, sequentially connecting each direction path according to the adjacent position relation, and extending the continuous range of the direction trend to obtain a direction track connection path set; s202, extracting direction information of adjacent pixels according to extending directions in the direction track connection path set, sequentially judging whether the peripheral direction trend continues to the original track, selecting image parts with consistent directions, and extending a direction channel range to obtain a direction channel extending area; s203, calling the image area in the direction channel extension area, extracting corresponding image content from the original image, supplementing the extension part to the corresponding position of the image, and integrally completing the connection processing of the direction track to obtain the direction path recombined image content.
5. The method for deep learning and classifying visible light image surface defects according to claim 1, wherein the step of obtaining the luminance jump boundary region map comprises the steps of: S301, referring to image areas in the direction path reorganization image content, extracting brightness information at each edge, detecting distribution differences of pixels in the edge direction, and distinguishing brightness trend among different areas according to brightness change ranges inside and outside the edge to obtain a brightness change trend set; S302, searching area boundaries in the brightness difference sets according to the distribution directions of the brightness differences in the brightness change trend sets, screening image ranges in which the brightness changes are continuously distributed, merging pixel ranges according to edge continuity, and obtaining edge continuous area segments; s303, extracting pixel data corresponding to the edges of the image range positioned in the edge continuous area segment, distributing the pixel data to a new processing image layer in an image space, forming an image presentation structure based on brightness jump positions, and obtaining a brightness jump boundary area diagram.
6. The method for deep learning and classifying visible light image surface defects according to claim 1, wherein the step of obtaining the color value corrected image frame comprises the steps of: s401, reading an image part extracted from the brightness jump boundary region diagram, sequentially extracting color distribution in a red channel, a green channel and a blue channel, detecting direction characteristics of each channel at adjacent positions, and performing alignment operation according to the extending direction among the color distribution to obtain a channel direction alignment image block; S402, calling the channel direction to align the color content in the image block, extracting the color extension directions of the red channel, the green channel and the blue channel in the image area, performing sequential adjustment on the color channels with direction offset, and performing unified processing according to the space trend of the color extension to obtain a color direction unified image segment; S403, according to the adjusted areas in the uniform graph segments of the color directions, correspondingly replacing the image parts at the same positions in the original graph, fusing the corrected color areas into the original image frame structure, outputting the complete image frame content, and obtaining the color value corrected image frame.
7. The method for deep learning classification of surface defects in visible light images according to claim 1, further comprising the step of S5: s5, extracting image content in the color value correction image frame, carding continuous pixel areas, extracting boundaries at positions with consistent color expression and direction trend, indicating areas of pixel areas in the boundary range, carrying out name assignment in combination with image characteristic information, and outputting visible light image surface defect classification results; The visible light image surface defect classification result comprises a defect type set, a defect area number and a defect grade label.
8. The method for deep learning and classifying visible light image surface defects according to claim 7, wherein the step of obtaining the visible light image surface defect classification result is as follows: s501, extracting image content in the color value correction image frame, screening continuous areas according to the position relation between adjacent pixels, carrying out area carding on the parts with consistent color change directions, and extracting edge boundaries according to color characteristics and space directions to obtain a color edge definition range; s502, detecting the morphological change and the color trend of pixels in each region according to the image parts defined in the color edge definition range, and screening the image parts with uniform trend as division basis by combining edge continuity and the structure direction in the region to obtain the image structure region distribution; s503, calling a texture arrangement mode and boundary characteristics in image content according to the position and the morphological characteristics of each region in the image structure region distribution, carrying out name assignment and region naming on the defect types corresponding to each region, and obtaining the classification content corresponding to the image to obtain the visible light image surface defect classification result.

Description

Deep learning classification method for visible light image surface defects Technical Field The invention relates to the technical field of image recognition, in particular to a visible light image surface defect deep learning classification method. Background The technical field of image recognition relates to automatic detection and classification of targets or features in images by using methods such as computer vision, deep learning and the like, and core matters comprise image acquisition, pretreatment, feature extraction, classification judgment and the like, and are widely applied to the fields of industrial detection, medical analysis, security monitoring and the like. The important point is to improve the accuracy and automation level of recognition, and often, effective processing of complex image information is realized by constructing a deep neural network model. The traditional visible light image surface defect deep learning classification method is a method for acquiring a workpiece surface image under the visible light condition and classifying surface defects in the image through a convolutional neural network and other models. The method generally takes a defect image with a label as a training sample, extracts characteristics through multi-layer rolling and pooling operation, outputs defect types by using a full-connection layer, optimizes model parameters by adopting a supervised learning mode, and is widely used for identifying defects such as scratches and cracks on the surfaces of materials such as metals and ceramics. The existing method is lack of modeling on the direction relation among pixels, and a microstructure mutation area cannot be identified, so that structural defects are easy to miss. The image break location is often judged as background, creating a feature loss. The boundary brightness change processing is rough, and the boundary fuzzy region classification is easy to be inaccurate. The lack of alignment mechanism of color information among channels is susceptible to local chromatic aberration, resulting in recognition errors. In images with various defect types and complex structures, classification performance is unstable, and defect judgment with clear boundaries and continuous structures is difficult to finish. Disclosure of Invention In order to solve the technical problems in the prior art, the embodiment of the invention provides a visible light image surface defect deep learning classification method. The technical scheme is as follows: The deep learning classification method for the visible light image surface defects comprises the following steps: S1, collecting red, green and blue channel information in a wind power blade image, carrying out brightness fusion on channel content, processing brightness directions between each pixel and adjacent pixels in the image, extracting position points in a direction mutation area, and keeping pixel distribution in the image to obtain a direction change breakpoint position diagram; S2, connecting the direction tracks in the image areas extracted from the direction change breakpoint position diagram, extending the direction information of adjacent pixels to enable the direction interval areas to be continuous, and taking the processed image as a subsequent input to obtain direction path recombined image content; S3, referring to image areas in the direction path reorganization image content, extracting edge brightness expression, observing brightness changes among different areas, separating out image parts with obvious brightness change trend, processing the change trend concentrated area into continuous image edges, and outputting a processing image layer to obtain a brightness jump boundary area diagram; S4, reading the image part extracted from the brightness jump boundary region image, respectively processing the color directions of red, green and blue channels, keeping the color relation of the region consistent by adjusting the color directions, replacing the processed image part with the corresponding region of the original image, and outputting a color value correction image frame; And S5, extracting image content in the color value correction image frame, carding continuous pixel areas, extracting boundaries at positions with consistent color expression and direction trend, indicating areas of pixel areas in the boundary range, carrying out name assignment in combination with image characteristic information, and outputting visible light image surface defect classification results. As a further scheme of the invention, the direction change breakpoint position map comprises breakpoint position coordinates, a breakpoint pixel point set and breakpoint distribution density, the direction path reorganization image content comprises reorganization direction channels, a direction continuity constraint domain and a fracture completion domain, the brightness jump boundary region map comprises jump boundary outlines, brigh