Ventilation equipment blade quality detection method based on image processing

Abstract

The invention discloses a ventilation equipment blade quality detection method based on image processing, relates to the technical field of image processing quality detection, and is used for solving the problem of inaccurate blade defect detection. According to the method, the original image of the blade is obtained by adaptively adjusting exposure time, industrial camera gain and a controllable light source combination, reflection and shadow interference is inhibited by adopting extreme region masking, multi-frame compensation and local brightness balance, a contour enhancement image and a texture enhancement image are generated, an outer contour and a skeleton are extracted on the basis of a design contour, and the texture enhancement image is obtained. Re-sampling the texture along the skeleton and the normal direction, constructing a blade surface expansion view and a quality analysis coordinate system, detecting multiple types of surface defects in an expansion domain, forming a blade quality feature vector by combining chord length, elongation, section width and bending deviation, outputting a quality grade and a disposal suggestion, and realizing unified quantification of geometric deviation and surface defects of the blade. And the online detection stability and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing quality inspection technology, and more specifically, to a method for inspecting the quality of ventilation equipment blades based on image processing. Background Technology

[0002] In the field of ventilation equipment manufacturing and assembly, components such as fan blades and air conditioning ventilation unit blades are widely used for air transport and flow field regulation. Their geometric accuracy and surface quality directly affect the overall efficiency, noise level, and long-term operational reliability. Currently, production lines generally employ manual visual inspection combined with sampling using measuring tools. Some companies have introduced simple industrial camera inspections, performing threshold segmentation and edge detection on individual blade images to determine if the blades are deformed or have obvious chipping or scratches. These methods largely rely on fixed exposure and fixed light source arrangements, primarily adjusting for specific angles of a particular blade model. Inspection results are usually based on manual interpretation of "pass" or "fail," making it difficult to establish unified, reusable quantitative quality indicators.

[0003] However, in actual production scenarios, ventilation equipment blades are often made of metal or have coated surfaces, exhibiting strong specular reflection and localized shadows. Furthermore, the blades possess a three-dimensional twisted structure, with significant differences in chord length, span, and leading and trailing edge shapes among different blade models. When imaging using single exposure and a fixed light source, the same image may simultaneously contain overexposed highlights and severely underexposed shadows, making it difficult for traditional algorithms based on global thresholds or simple edge operators to reliably extract the true outer contour and fine surface defects. Simultaneously, directly determining defects in two-dimensional pixel coordinates cannot accurately locate cracks, notches, pits, and other defects to their specific positions along the blade's length and leading and trailing edges. It is even more difficult to uniformly express geometric deviations and various surface defects as structured quality characteristics, resulting in a lack of comparability and traceability between different batches and models of test results. This hinders subsequent process optimization and quality statistical analysis. Therefore, in the aforementioned scenarios, achieving stable detection and unified quantification of blade geometric contours and various types of defects under conditions of strong reflection, shadows, and three-dimensional curved blade morphology constitutes a significant technical challenge in existing technologies. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the following solution is proposed to address the problem of inaccurate blade defect detection in the aforementioned background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The image processing-based method for detecting the quality of ventilation equipment blades includes the following steps: At the inspection station, an industrial camera and a controllable light source are used to acquire images of the blades. The exposure and light source are adjusted according to the brightness and saturation areas of the image to obtain the original image of the blades. The original leaf image is preprocessed for illumination and reflection, highlight and shadow areas are marked, local brightness equalization is applied to the remaining areas, and multi-frame information is used to compensate for the highlight and shadow areas to generate contour enhancement map and texture enhancement map. Using the contour enhancement map as input, the blade region is located and the outer contour is extracted. The broken contour is aligned and completed by combining the model design contour. The skeleton along the length direction is extracted within the outer contour. Using the skeleton as the vertical reference, the blade region in the texture enhancement map is resampled along the skeleton and normal directions to construct the blade surface unfolding map and establish a quality analysis coordinate system with the skeleton position and the distance to the leading and trailing edges as coordinates. In the unfolded image of the blade surface, response operators along the skeleton and normal, as well as gray-scale operators, are used to extract candidate regions for cracks, notches, pits, and coating defects, resulting in multiple types of defect regions. The geometric parameters obtained from the skeleton and outer contour, as well as the type, size, and position of various defect regions in the quality analysis coordinate system, are combined to form the blade quality feature vector, and the detection results corresponding to the blade quality feature vector are output.

[0006] Furthermore, the steps of acquiring blade images at the inspection station using an industrial camera and a controllable light source, adjusting the exposure and light source according to the brightness and saturation areas of the image, and obtaining the original blade image include: The brightness histogram, high-brightness saturated pixel ratio, and low-brightness pixel ratio are calculated from the preview image captured when the blade enters the inspection station. The obtained statistics are matched with the preset imaging parameter configuration table. The corresponding exposure time, industrial camera gain, and controllable light source combination are selected from the imaging parameter configuration table. The parameters of the industrial camera and controllable light source are set. After the parameter settings are completed, the original image of the blade is captured.

[0007] Furthermore, the steps of performing illumination and reflection preprocessing on the original leaf image, marking highlight and shadow areas, local brightness equalization of the remaining areas, and generating contour enhancement and texture enhancement maps include: Highlight and shadow areas are detected based on pixel brightness threshold and local gradient threshold in the original image of the leaf, and extreme region masks are generated according to the connected regions. Pixel areas outside the extreme region mask are regarded as non-extreme regions. Local brightness equalization and local contrast stretching are performed on non-extreme regions. The images of non-extreme regions after local brightness equalization and local contrast stretching are used as the base images for contour enhancement and texture enhancement.

[0008] Furthermore, the steps of compensating for highlight and shadow areas using multi-frame information include: While keeping the position of the industrial camera unchanged, multiple frames of original images of the blades are acquired by sequentially switching between different polarization directions or different combinations of light sources. Spatial registration is performed on multiple frames of original leaf images. In the highlight areas marked by the mask in extreme regions, the pixels with the lowest saturation or brightness within a preset range are selected to replace the corresponding pixels. Interpolation compensation is performed within the shadow area marked by the mask in the extreme region using pixels from adjacent non-extreme regions and texture pixels at the corresponding positions of the previous batch of qualified blades to obtain the compensated original image of the blade used to generate the contour enhancement map and texture enhancement map.

[0009] Furthermore, the steps of taking the contour enhancement map as input, locating the blade region and extracting the outer contour, aligning and completing the broken contour in conjunction with the model design contour, and extracting the skeleton along the length direction within the outer contour include: In the contour enhancement map, the initial position and size of the blade region are obtained through target detection or template matching. In the blade region, edge detection and threshold segmentation are used to extract the initial outer contour. The initial outer contour is aligned with the design contour corresponding to the blade model and the scale is normalized. At the location where there is a break in the initial outer contour, the contour fragments at the corresponding positions of the design contour are used to complete and smooth it. In the interior of the completed outer contour, the blade skeleton distributed along the blade length direction is extracted based on distance transformation and thinning algorithms.

[0010] Furthermore, the steps of using the skeleton as the longitudinal reference, resampling the blade region in the texture enhancement map along the skeleton and normal directions, constructing the blade surface unfolding map, and establishing a quality analysis coordinate system with the skeleton position and the distance to the leading and trailing edges as coordinates include: In the texture enhancement map, the blade region is divided into several sampling positions along the skeleton direction using the blade skeleton. At each sampling position, the normal direction sampling line from the leading edge to the trailing edge is determined according to the outer contour. The pixel arrays arranged in the longitudinal and lateral positions are obtained by resampling along the sampling line at a preset interval. The pixel arrays are used to form a unfolded map of the blade surface. The quality analysis coordinate system of the blade surface is established by using the sampling index along the blade skeleton as the longitudinal coordinate and the normalized distance from the blade skeleton to the leading edge and from the blade skeleton to the trailing edge as the lateral coordinate.

[0011] Furthermore, the steps for extracting candidate regions for cracks, notches, pits, and coating defects in the blade surface development diagram using response operators along the skeleton and normal, as well as grayscale calculation operators, to obtain multi-type defect regions include: In the unfolded image of the blade surface, a linear response operator is applied along the longitudinal direction to extract slender strip-shaped high-response regions as candidate regions for cracks and scratches. An edge response operator is applied along the transverse direction to extract edge morphology change regions near the leading and trailing edges as candidate regions for notches. The gray-level mean and gray-level variance are calculated within a local window and compared with adjacent windows. Isolated regions with gray-level anomalies are extracted as candidate regions for pitting and coating defects. The candidate regions are clustered according to connected domains and the region boundaries are recorded to form a set of multi-type defect regions.

[0012] Furthermore, after obtaining the multi-type defect regions, the steps of classifying and extracting parameters from these regions include: Based on the boundary of each multi-type defect region, defect image blocks are cropped from the unfolded image of the blade surface. Size normalization and grayscale normalization are performed on the defect image blocks. The normalized defect image blocks are then input into a pre-trained deep learning classification model. The deep learning classification model outputs defect type labels and defect contour masks. Based on the defect contour masks, the area, length, width, and principal axis angle of the defect region are calculated. The defect type labels, area, length, width, and principal axis angle are recorded in a multi-type defect region set.

[0013] Furthermore, the step of combining the geometric parameters obtained from the skeleton and outer contour, as well as the type, size, and position of various defect regions in the quality analysis coordinate system, into a blade quality feature vector includes: The chord length, span, cross-sectional width, bending deviation, and shape parameters of the skeleton along the length direction are combined into a geometric parameter subvector; The defect type label, defect area, defect length, defect width, defect principal axis direction angle, and longitudinal and transverse coordinates in the quality analysis coordinate system of each multi-type defect region are combined to form a defect parameter sub-vector. All defect parameter subvectors of the same blade are concatenated or compressed in a preset order to obtain a defect summary subvector. The geometric parameter subvectors are then concatenated with the defect summary subvector to form the blade quality feature vector.

[0014] Furthermore, the detection results corresponding to the output blade quality feature vector include: Input the blade quality feature vector into the quality assessment rule base or quality assessment model, and determine the blade quality grade and corresponding treatment recommendations based on the value range of bending deviation, cross-sectional width and span in the geometric parameter sub-vector and the combination relationship of defect type, defect size and defect location distribution in the defect summary sub-vector. The blade identification information, blade quality feature vector, quality grade, and treatment recommendations are organized into a structured record and written into the quality database. The test results, which include the quality grade, geometric parameter summary, and defect list, are output on the test terminal.

[0015] The technical effects and advantages of the image processing-based method for detecting the quality of ventilation equipment blades in this invention are as follows: This invention uses a preview imaging statistics and imaging parameter configuration table to adaptively adjust the exposure time, industrial camera gain, and controllable light source combination. It also combines extreme area masking, multi-frame compensation, and local brightness equalization to reduce the impact of strong reflections and shadows on the image quality of the blades, and stably obtain contour enhancement maps and texture enhancement maps under complex lighting conditions. Based on the blade surface development diagram and quality analysis coordinate system constructed from the model design outline and refined skeleton, geometric measurement and defect location of three-dimensional curved blades are realized in a unified two-dimensional coordinate system. This makes the results of different batches and different inspection stations comparable and traceable. Then, by extracting multiple types of candidate regions for cracks, notches, pits and coating defects on the development diagram, and organizing the geometric parameters and defect parameters into a blade quality feature vector, the system outputs quality level and treatment suggestions in conjunction with a quality assessment rule base or quality assessment model. This achieves unified quantitative evaluation of multiple types of defects and geometric deviations, improves the defect detection rate and judgment consistency, reduces manual dependence, and is suitable for large-scale online quality inspection of ventilation equipment blades. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of the image processing-based method for detecting the quality of ventilation equipment blades according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In order to achieve the above objectives, Figure 1 A schematic diagram of the image processing-based method for detecting the quality of ventilation equipment blades according to the present invention is provided, which specifically includes the following steps; At the inspection station, an industrial camera and a controllable light source are used to acquire images of the blades. The exposure and light source are adjusted according to the brightness and saturation areas of the image to obtain the original image of the blades. The original leaf image is preprocessed for illumination and reflection, highlight and shadow areas are marked, local brightness equalization is applied to the remaining areas, and multi-frame information is used to compensate for the highlight and shadow areas to generate contour enhancement map and texture enhancement map. Using the contour enhancement map as input, the blade region is located and the outer contour is extracted. The broken contour is aligned and completed by combining the model design contour. The skeleton along the length direction is extracted within the outer contour. Using the skeleton as the vertical reference, the blade region in the texture enhancement map is resampled along the skeleton and normal directions to construct the blade surface unfolding map and establish a quality analysis coordinate system with the skeleton position and the distance to the leading and trailing edges as coordinates. In the unfolded image of the blade surface, response operators along the skeleton and normal, as well as gray-scale operators, are used to extract candidate regions for cracks, notches, pits, and coating defects, resulting in multiple types of defect regions. The geometric parameters obtained from the skeleton and outer contour, as well as the type, size, and position of various defect regions in the quality analysis coordinate system, are combined to form the blade quality feature vector, and the detection results corresponding to the blade quality feature vector are output.

[0019] Step 1: At the inspection station, an industrial camera and a controllable light source are used to acquire images of the blades. The exposure and light source are adjusted according to the brightness and saturation areas of the image to obtain the original image of the blades. The specific implementation is as follows: The process of acquiring images of the ventilation equipment blades using an industrial camera and a controllable light source at the inspection station, adjusting the exposure and light source according to the brightness and saturation areas of the image, and obtaining the original image of the blades is preferably implemented as follows: An industrial camera and a controllable light source are fixedly installed at the inspection station. The imaging field of view of the industrial camera covers the inspection area traversed by the ventilation equipment blades. The controllable light source is positioned above or to the sides of the blades to provide stable illumination. When the ventilation equipment blades enter the inspection station via the conveyor mechanism, the external control system triggers the industrial camera to acquire one or more preview frames. The preview frames refer to the blade images acquired using the current default imaging parameters before the final determination of the exposure time, industrial camera gain, and controllable light source combination. These preview frames are used for statistical analysis of the current ambient brightness and reflectivity, rather than being directly used as the original images of the blades for subsequent quality inspection.

[0020] In order to extract statistics that can be used to adjust imaging parameters from the preview image, the system counts the brightness value of each pixel in the preview image, divides the brightness value into a preset number of brightness levels from the lowest brightness level to the highest brightness level, and counts the number of pixels falling into each brightness level to obtain a brightness histogram.

[0021] The brightness histogram represents the number of pixels corresponding to each brightness level, and is used to reflect the overall brightness distribution of the preview image.

[0022] Based on this, a brightness threshold for high-brightness saturated pixels and a brightness threshold for low-brightness pixels are preset. The brightness threshold for high-brightness saturated pixels is used to distinguish pixels that are close to the highest brightness that an industrial camera can represent or that have reached saturation. The brightness threshold for low-brightness pixels is used to distinguish pixels that are close to dark areas.

[0023] The number of pixels with brightness values ​​higher than the high-brightness saturation pixel brightness threshold is counted, and the ratio of this number to the total number of pixels in the preview image is calculated. This is expressed as follows: the high-brightness saturation pixel count is divided by the total number of pixels to obtain the high-brightness saturation pixel percentage. The number of pixels with brightness values ​​lower than the low-brightness pixel brightness threshold is counted, and the low-brightness pixel percentage is divided by the total number of pixels to obtain the low-brightness pixel percentage.

[0024] The parameters of high-brightness saturated pixel ratio and low-brightness pixel ratio are used to characterize the degree of overexposure and underexposure of the current image.

[0025] The brightness thresholds for high-brightness saturated pixels and low-brightness pixels were determined by engineers through multiple experiments during the system debugging phase. The specific method was as follows: Multiple sets of images were acquired under different exposure times, industrial camera gains, and controllable light source combinations. The range of the brightest and darkest pixels that could be tolerated without losing leaf details was observed. The corresponding brightness levels were used as the brightness thresholds for high-brightness saturated pixels and low-brightness pixels, and were used as fixed parameters for operation.

[0026] The brightness histogram, the proportion of high-brightness saturated pixels, and the proportion of low-brightness pixels are converted into specific imaging parameters for industrial cameras and controllable light sources, and an imaging parameter configuration table is established.

[0027] The imaging parameter configuration table is a lookup table indexed by statistical ranges and containing combinations of imaging parameters. Each record contains a brightness histogram feature range, a set of high-brightness saturated pixel ratio ranges, a set of low-brightness pixel ratio ranges, and the exposure time, industrial camera gain, and controllable light source combination that match the range combination.

[0028] The feature range of the brightness histogram can be described by extracting the pixel proportion of several representative brightness level segments in the brightness histogram. For example, the pixel proportion of the low brightness segment, the medium brightness segment, and the high brightness segment in the brightness histogram can be selected as the brightness histogram features. The proportion range of each segment is determined by statistical analysis of sample images of different ambient light and different blade surface conditions during the debugging phase.

[0029] The range of high-brightness saturated pixel ratio and low-brightness pixel ratio were also obtained through experiments during the debugging phase. During the experiment, the exposure time, industrial camera gain and controllable light source combination were adjusted to observe the value range of the corresponding high-brightness saturated pixel ratio and low-brightness pixel ratio under the condition that the key features of the blade are clearly visible and there is no overexposure or severe underexposure. The value range, together with the corresponding exposure time, industrial camera gain and controllable light source combination, were recorded in the imaging parameter configuration table.

[0030] In actual operation, for each blade under test, the brightness histogram, the proportion of high-brightness saturated pixels, and the proportion of low-brightness pixels are first obtained using the preview image. These three statistics are then matched with each record in the imaging parameter configuration table. The matching method can be as follows: In the imaging parameter configuration table, find records whose brightness histogram feature range includes the current brightness histogram feature, whose high brightness saturation pixel ratio range includes the current high brightness saturation pixel ratio, and whose low brightness pixel ratio range includes the current low brightness pixel ratio. If multiple records meet the conditions, select the record with the highest priority according to the priority field. If there is no record that matches exactly, select the record with the smallest difference from the current statistic.

[0031] The system reads the exposure time, industrial camera gain, and controllable light source combination from the selected record. It sets the exposure time and gain parameters to the industrial camera's default settings and opens or closes the corresponding controllable light source channel according to the definition of the controllable light source combination. After setting the exposure time, industrial camera gain, and controllable light source combination, the control system triggers the industrial camera to acquire images again, obtaining the original image of the blade for subsequent illumination and reflection preprocessing and quality inspection.

[0032] Through the above steps, the acquisition process of the original leaf image is entirely driven by the brightness histogram, the proportion of high-brightness saturated pixels, and the proportion of low-brightness pixels. This achieves adaptive imaging for different ambient lighting conditions and leaf surface reflectivity, ensuring that the original leaf image remains within the expected grayscale dynamic range in most pixel areas. Step 2: Perform illumination and reflection preprocessing on the original leaf image, mark highlight and shadow areas, perform local brightness equalization on the remaining areas, compensate for highlight and shadow areas using multi-frame information, and generate contour enhancement and texture enhancement maps. The specific implementation is as follows: The preferred process for preprocessing the original leaf image by illumination and reflection, marking highlight and shadow areas, performing local brightness equalization on the remaining areas, compensating for highlight and shadow areas with multi-frame information, and generating contour enhancement and texture enhancement maps includes: First, using the original image of the leaf as input, the brightness value and local gradient value are calculated for each pixel in the original image. The brightness value can be represented by the grayscale value of a single channel in a grayscale image, or by performing a color space conversion on a color image and taking the component representing brightness after the conversion as the brightness value. The local gradient value is obtained by establishing a local window of a preset size at each pixel location and calculating the difference between the maximum and minimum brightness values ​​within that local window, which is used to characterize the degree of brightness change in the neighborhood of that pixel.

[0033] Pre-set brightness thresholds for highlight areas, shadow areas, and local gradient areas. The highlight area brightness threshold determines whether the pixel brightness is close to the upper limit of the industrial camera's brightness; the shadow area brightness threshold determines whether the pixel brightness is in a near-dark area; and the local gradient threshold determines whether there are significant brightness changes in a local area. These thresholds are determined statistically from multiple sets of leaf sample images during the system debugging phase. Specifically: Leaf images were acquired under different lighting conditions and imaging parameters. The pixel brightness and local gradient distribution range of manually determined high reflective areas and deep shadow areas were observed. The boundary values ​​that could stably distinguish normal areas within these ranges were used as thresholds. Pixels are marked as highlight areas if their brightness value is higher than the highlight area brightness threshold or their local gradient value is higher than the local gradient threshold. Pixels are marked as shadow areas if their brightness value is lower than the shadow area brightness threshold and their local gradient value is lower than the local gradient threshold. All pixels marked as highlight and shadow areas are aggregated into several extreme regions through connected component analysis, and extreme region masks are generated accordingly. Pixel regions not covered by the extreme region masks are uniformly defined as non-extreme regions.

[0034] For non-extreme areas, in order to reduce the impact of uneven ambient lighting on subsequent segmentation and defect detection, local brightness equalization and local contrast stretching are performed on non-extreme areas.

[0035] Local brightness equalization refers to creating a local window at each pixel location, statistically analyzing the distribution of pixel brightness within that local window, and remapping the brightness values ​​within that local window to a preset local brightness range. This allows for the appropriate enhancement of areas with low brightness and the appropriate compression of areas with high brightness, thereby reducing the overall brightness difference between different locations while maintaining the edge structure.

[0036] Local contrast stretching expands the effective range of brightness value distribution within each local window on the basis of local brightness equalization, mapping the relatively concentrated brightness range within the local window to a wider brightness range, making the grayscale changes in the local area more obvious.

[0037] The size of the local window is preset based on the blade width and surface texture scale. For example, a rectangular window containing a certain number of pixels along the blade length direction and a certain number of pixels along the blade width direction can be selected. The window size is selected during the system debugging phase by comparing the degree of outline sharpness and texture preservation under different window sizes. Local brightness equalization and local contrast stretching are only performed in non-extreme areas. Pixels in extreme areas retain their original brightness values ​​during this stage. After processing, a non-extreme region image with local brightness equalization and local contrast stretching is obtained. This non-extreme region image serves as the base image for subsequent generation of contour enhancement map and texture enhancement map.

[0038] To reliably compensate for highlights and shadows, multiple frames of original blade images are used. Specifically, while keeping the industrial camera position constant, a controllable light source is sequentially switched between different polarization directions or different light source combinations to continuously acquire multiple frames of original blade images for the same blade within a short period of time. The original blade images are basically consistent in blade shape and position, with differences only in reflective texture and shadow distribution.

[0039] Spatial registration is performed on multiple frames of original blade images by feature point matching or template-based alignment. Spatial registration refers to calculating the translation and rotation relationships between each frame of images and mapping the multiple frames of original blade images to a unified coordinate system, so that the pixel positions of the blade surface at the same physical location are aligned in the multiple frames of images.

[0040] After spatial registration, the system uses the extreme region mask obtained from the single-frame original leaf image to extract the pixel brightness values ​​at corresponding positions in multiple frames of the original leaf image. For the highlight areas marked by the extreme region mask, the system compares the pixel brightness and saturation at that position in each frame, and selects pixels whose brightness is within the preset normal brightness range and has not reached the upper limit of brightness as candidate pixels. If the position is saturated in multiple frames, the frame with the lowest brightness is selected as the compensation value, and this compensation value replaces the highlight pixel brightness at the corresponding position in the original leaf image. For the shadow areas marked by the extreme region mask, the system first spatially searches for non-extreme region pixels adjacent to the shadow position, and then extracts the statistical characteristics of brightness from these non-extreme region pixels, specifically: Within the neighborhood, the brightness values ​​of multiple non-extreme region pixels are collected, and the average or median of these brightness values ​​is calculated. This statistical value is used as the neighborhood brightness reference. Simultaneously, the texture brightness features of the previous batch of qualified leaves at the corresponding position are searched in the quality database. The neighborhood brightness reference is multiplied by a first scaling factor, and the historical texture brightness reference is multiplied by a second scaling factor. The products of the two are summed and then divided by the sum of the first and second scaling factors to obtain the target brightness value for shadow area brightness compensation, i.e., a single compensation brightness value. This compensation brightness value is used to replace the original brightness of the shadow area pixels.

[0041] After the above compensation process, a compensated original image of the leaf is obtained, in which the brightness of the highlight and shadow areas is corrected.

[0042] Based on the original image of the blade after compensation, contour enhancement maps and texture enhancement maps are generated by combining the results of local brightness equalization and local contrast stretching in the aforementioned non-extreme regions. The specific method is as follows: In non-extreme regions, the pixel brightness values ​​after local brightness equalization and local contrast stretching are directly used. In extreme regions, the compensated brightness values ​​in the original image of the blade are used. The two types of pixels are seamlessly stitched together using an extreme region mask to obtain an intermediate image with a more uniform overall brightness distribution. Subsequently, based on the intermediate image, two different post-processing processes were executed. The post-processing process for generating the contour enhancement map enhanced the edge structure of the blade's outer contour and leading and trailing edges by smoothing fine texture noise and preserving edges with abrupt brightness changes. Specifically, while smoothing the brightness within a local window, the original brightness difference was preserved at positions where the brightness change exceeded a preset edge threshold. The post-processing process for generating the texture enhancement map highlighted the texture details and minor defects on the blade surface by enhancing local contrast and fine brightness changes. Specifically, a high-frequency component enhancement process was performed again on the intermediate image, and the high-frequency component was superimposed back onto the intermediate image to obtain the texture enhancement result.

[0043] After the above processing, the resulting contour enhancement map is mainly used for subsequent blade region localization and outer contour extraction, and the resulting texture enhancement map is mainly used for subsequent defect candidate region extraction and defect recognition. This achieves illumination and reflection preprocessing, extreme region marking, multi-frame information compensation, and dual-channel enhancement map generation for the original blade image.

[0044] Step 3: Using the contour enhancement map as input, locate the blade region and extract the outer contour. Combine the model design contour to align and complete the broken contour. Extract the skeleton along the length direction within the outer contour. Specifically, the implementation is as follows: Using the contour enhancement image as input, the first step is to locate the leaf region within the entire image. The contour enhancement image is a grayscale image that highlights the grayscale variations at the outer edge of the leaf after illumination preprocessing and local contrast processing.

[0045] During the initialization phase, engineers provide a blade template image or target feature information of the blade outline based on the approximate shape of the ventilation equipment blade and its field of view position in the inspection station. The target feature information may include the aspect ratio of the blade in the image, the direction of the long side of the circumscribed rectangle, and the approximate orientation of the blade root position.

[0046] In actual operation, object detection or template matching is used. Candidate detection windows slide on the contour enhancement map, and the similarity between each candidate detection window and the template image is calculated. The similarity is achieved by comparing the edge intensity distribution within the candidate detection window with the edge intensity distribution of the template. When the similarity is higher than a preset similarity threshold, the rectangular region corresponding to the candidate detection window is marked as a leaf region. The similarity threshold was determined statistically from multiple sets of sample images containing and not containing leaves during the system debugging phase, ensuring that regions containing leaves can be distinguished from background regions in most cases.

[0047] After the localization process, one or more rectangular blade regions are obtained, and the outer contour of each blade region is extracted.

[0048] Within the blade region, to extract the outer contour, the grayscale values ​​of the blade region in the contour enhancement image are first binary segmented to separate the blade body from the background. Specifically, this is implemented as follows: The gray-level distribution within the leaf area is statistically analyzed. A binary segmentation gray-level threshold is selected based on the gray-level histogram. Pixels with gray-level values ​​higher than the binary segmentation gray-level threshold are classified as leaf candidate pixels, and pixels with gray-level values ​​lower than the binary segmentation gray-level threshold are classified as background candidate pixels. The grayscale threshold for binary segmentation is determined by analyzing the grayscale histograms of multiple frames of leaf region images during the debugging process. This ensures that most of the outer contour neighborhood pixels in the leaf region are classified into the leaf candidate pixel category. After obtaining the leaf candidate pixels, isolated connected regions with areas smaller than the preset area threshold are removed by connected component analysis. These isolated connected regions are treated as noise and deleted. Only the largest connected region with an area greater than the area threshold is retained as the main body region of the leaf. Subsequently, edge detection is performed on the boundary of the blade body area. Edge detection adopts the method of calculating gray-level gradient around the blade body area. Contour points are marked at the locations where the gradient value is greater than the contour strength threshold. The contour strength threshold is determined by engineers by observing that the contour is clear and does not introduce too many false edges. All contour points are sorted in row and column order and connected by multiple polylines to obtain the initial outer contour. Due to the influence of strong reflection and local shading, the initial outer contour may have local breaks or offsets. Therefore, a design contour corresponding to the blade model is introduced. The design profile is a theoretical profile generated by a three-dimensional computer-aided design model during the design phase of ventilation equipment blades. By calibrating an industrial camera at the inspection station, the design profile is converted from physical coordinates to image coordinates, resulting in a set of design profile points in the same image coordinate system as the profile enhancement image.

[0049] Alignment is achieved by aligning the initial outer contour point set and the design contour point set. Specifically, within the allowable translation and scaling range, a set of translation and scaling parameters is searched to maximize the overall overlap between the two sets. The optimal alignment is determined by calculating the translation and scaling parameters that minimize the average distance between the two point sets. After coordinate transformation of the initial outer contour point set according to these parameters, it is aligned with the design contour point set. In areas where the initial outer contour has breaks, missing contour points are inserted along the corresponding curve segments of the design contour, and the joints are smoothed to obtain a continuous, smooth, and highly consistent complete outer contour that closely matches the actual blade contour.

[0050] After completing the outer contour extraction, the blade skeleton distributed along the blade length direction is further extracted inside the outer contour. Specifically, this is achieved as follows: The main region of the blade within the completed outer contour is converted into a binary mask image. Pixels inside the outer contour are marked as internal pixels of the blade, and pixels outside the outer contour are marked as background pixels. In the binary mask image, a thinning operation is performed on the internal pixels of the blade. The thinning operation means that, under the premise of ensuring that the overall topological structure is not broken, the internal pixels of the blade that meet the preset deletion conditions are gradually deleted from the boundary to the inside, and only the pixels representing the central structure of the region are retained. Before refining, the distance distribution from the internal pixels of each leaf to the outer contour can be calculated based on the distance transformation to help determine the central region of the skeleton and further improve the skeleton connectivity and centering.

[0051] The deletion criteria can be determined using a local neighborhood model. For example, it can be determined whether the pixel is located at a local edge in the neighborhood of the current pixel and whether deletion will not create a new disconnected region. After iteratively performing the thinning operation until no pixel meets the deletion criteria, the remaining internal pixels of the leaf form a long and thin pixel chain extending from the leaf root to the leaf tip. This long and thin pixel chain is the leaf skeleton. Based on the major axis direction of the blade in the imaging coordinates, the skeleton pixels are sorted according to their positions along the major axis direction to generate an ordered sequence of blade skeleton points. Using the calibration information of the industrial camera, the image coordinates of each blade skeleton point are converted into physical coordinates to obtain a sequence of skeleton physical coordinates along the blade length direction. This blade skeleton serves as both a longitudinal reference for subsequent blade surface resampling and unfolding and a means to calculate geometric parameters along the blade length direction, such as bending tendency and local offset, based on the skeleton physical coordinate sequence.

[0052] In summary, the above steps enable the process of locating the blade region, extracting the outer contour, completing the breakage repair based on the design contour, and extracting the blade skeleton along the length direction, using the contour enhancement map as input.

[0053] Step 4: Using the skeleton as the vertical reference, resample the blade region in the texture enhancement map along the skeleton and normal directions to construct the blade surface unfolding map and establish a quality analysis coordinate system with the skeleton position and the distance to the leading and trailing edges as coordinates. The specific implementation is as follows: Using the blade skeleton as the longitudinal reference, the blade region in the texture enhancement map is resampled along the skeleton and normal directions, and a blade surface unfolding map and a quality analysis coordinate system for the blade surface with the skeleton position and the distance to the leading and trailing edges as coordinates are constructed. Preferably, this is implemented in the following manner: First, based on the obtained complete outer contour and blade skeleton, the complete outer contour includes the leading edge contour line and the trailing edge contour line. The blade skeleton is composed of a series of skeleton points distributed along the blade length direction. Each skeleton point has horizontal and vertical coordinates in the image coordinates, and the corresponding physical coordinates can be obtained through the calibration relationship. Longitudinal sampling is performed along the blade skeleton at a preset step size. The preset step size can be a certain proportion of the total length of the blade skeleton or a fixed number of pixels. During the engineering debugging phase, it is determined by comparing the smoothness of the unfolded image under different sampling densities and the computational load. For example, in one application, the blade skeleton can be divided into one hundred longitudinal sampling positions, each corresponding to a cross section along the blade length direction. For each longitudinal sampling position, based on the skeleton point and outer contour at that position, the shortest connection from the skeleton point to the leading edge contour line and the shortest connection from the skeleton point to the trailing edge contour line are calculated. The directions of the two shortest connections are basically perpendicular to the local tangent direction of the blade skeleton. These two connections together form the normal direction sampling line from the leading edge through the skeleton to the trailing edge in the image coordinates.

[0054] Record the starting point of each normal direction sampling line as the leading edge intersection point and the ending point as the trailing edge intersection point, and calculate the total distance from the leading edge intersection point to the trailing edge intersection point. This total distance is used as the total length of the normal direction at the longitudinal sampling position for normalization processing.

[0055] After obtaining the normal direction sampling line corresponding to each longitudinal sampling position, in order to map the blade region in the texture enhancement map to the regular blade surface unfolding map, resampling is performed on each normal direction sampling line according to a preset lateral sampling interval. The lateral sampling interval can be determined by dividing the total distance from the leading edge to the trailing edge into several equally spaced sampling points. For example, in the example of one hundred longitudinal sampling positions mentioned above, fifty lateral sampling points can be divided on each normal direction sampling line to obtain a blade surface unfolding map with one hundred rows and fifty columns. For each normal direction sampling line, the system determines the position of each lateral sampling point in sequence from the leading edge to the trailing edge. The physical position of each lateral sampling point is determined by its distance from the intersection point of the leading edge in the normal direction. This distance can be obtained by dividing the total distance from the leading edge to the trailing edge by the number of lateral sampling points to obtain the single-step distance, and then accumulating the single-step distances as integer multiples to obtain the position of each sampling point. Each lateral sampling point in the image coordinate system generally does not fall exactly on an integer pixel. The system uses interpolation to obtain the grayscale value of that location from the texture enhancement image. The interpolation method can select the grayscale values ​​of several neighboring pixels around that location, and perform a weighted average based on distance, calculating the grayscale value of the lateral sampling point based on the principle that the nearest neighbor pixel has a greater influence. The grayscale values ​​of all vertical sampling positions and their corresponding lateral sampling points are arranged according to the vertical sampling position as the row index and the lateral sampling point number as the column index, thus obtaining a regular matrix form of the blade surface unfolded image. In this unfolded image, the vertical direction from top to bottom corresponds to the order of the blade from the root to the tip, and the horizontal direction from left to right corresponds to the direction of the blade from the leading edge to the trailing edge. In this way, the blade surface, which was originally distributed as a curved surface in the texture enhancement image, is unfolded into a regular two-dimensional grayscale image, which facilitates subsequent unified processing.

[0056] While constructing the unfolded diagram of the blade surface, a mass analysis coordinate system for the blade surface is established to describe the position of any point on the blade within the unfolded domain. The mass analysis coordinate system for the blade surface uses two coordinate components: a longitudinal coordinate and a transverse coordinate. The longitudinal coordinate is used to represent the position along the blade skeleton direction and can be implemented in one of two ways: One approach is to directly use the serial number of the longitudinal sampling position. For example, starting from the end closest to the leaf root, we can number the first longitudinal position, the second longitudinal position, and so on until we reach the end closest to the leaf tip. In this way, the longitudinal coordinates are discrete integers. Another approach is to use the physical length of the leaf skeleton as a reference, and divide the physical length of each longitudinal sampling position from the leaf root by the total length of the leaf skeleton to obtain a longitudinally normalized position parameter between zero and one. The lateral coordinate is used to represent the distance relationship from the blade skeleton to the leading and trailing edges. The normalized distance from the leading edge to the trailing edge is preferred as the lateral coordinate. That is, the distance of each lateral sampling point along the normal direction from the intersection point of the leading edge is divided by the total distance along the normal direction from the intersection point of the leading edge to the intersection point of the trailing edge, resulting in a lateral position parameter between zero and one. Zero corresponds to the leading edge position, one corresponds to the trailing edge position, and the intermediate value corresponds to various positions inside the blade cross section.

[0057] This ensures that any pixel in the unfolded image of the blade surface not only has a row and column index position, but also corresponds to a vertical coordinate and a horizontal coordinate. The vertical coordinate reflects its position in the blade length direction, and the horizontal coordinate reflects its front and rear position on the cross section.

[0058] Furthermore, based on the already determined image coordinates of each sampling point in the original image and the physical coordinates obtained through calibration relationships, each coordinate pair in the quality analysis coordinate system of the blade surface can also correspond to a three-dimensional position in physical space. For example, in a simple application scenario, engineers can specify a strip area with a longitudinal coordinate within a certain range and a lateral coordinate close to the leading edge as a key area of ​​focus to check for pitting or coating defects on the blade leading edge within a specific length range, without having to return to the original image for complex geometric positioning, thus significantly reducing the complexity of defect location and statistical analysis.

[0059] Step 5: In the unfolded image of the blade surface, response operators along the skeleton and normal, as well as grayscale operators, are used to extract candidate regions for cracks, notches, pits, and coating defects, resulting in multiple types of defect regions. The specific implementation is as follows: After constructing the blade surface development diagram and the coordinate system for quality analysis of the blade surface, the system uses the blade surface development diagram as input and extracts candidate defect regions from the diagram according to the longitudinal coordinates along the blade skeleton direction and the lateral coordinates from the leading edge to the trailing edge. Specifically, the index ranges for the longitudinal and lateral directions are first determined in the blade surface development diagram. The longitudinal direction is numbered sequentially from the starting position of the longitudinal coordinates corresponding to the blade root to the ending position of the longitudinal coordinates corresponding to the blade tip, and the lateral direction is numbered sequentially from the starting position of the lateral coordinates corresponding to the leading edge to the ending position of the lateral coordinates corresponding to the trailing edge. Each pixel in the unfolded image of the blade surface has vertical and horizontal coordinates. The system applies different types of response operators and grayscale operators in the vertical and horizontal directions respectively to identify candidate areas for cracks, scratches, notches, pits and coating defects.

[0060] For cracks and scratches distributed along the blade length, a linear response operator along the longitudinal direction is used for detection. The linear response operator defines a pixel sequence along the longitudinal direction in the unfolded image of the blade surface, takes a continuous segment of pixels in the pixel sequence as an observation window, compares the gray value difference between the central pixel and several pixels before and after it in the observation window, and calculates the continuity of the gray value difference in the longitudinal direction. If a feature with a gray value significantly lower than the surrounding background gray value or a gray value change in the form of a thin strip appears in a continuous longitudinal position, the corresponding pixel is marked as a crack and scratch response point.

[0061] The specific parameters of the linear response operator, such as the longitudinal length of the observation window, the gray-scale difference judgment threshold, and the continuous length judgment threshold, were determined through repeated experiments on the surface development diagram of the blade with labeled crack samples during the debugging phase, so that the linear response operator can stably respond to crack and scratch morphology under different blade models and different surface textures. For notches located near the leading and trailing edge contours, an edge response operator along the lateral direction is used for detection. The edge response operator establishes a lateral observation window in the lateral direction. At each fixed longitudinal coordinate position, the lateral observation window slides point by point from the leading edge to the trailing edge, calculating the gray-level difference and gradient direction at the front and rear positions within the window. If a sharp change in gray-level occurs near the leading or trailing edge and the gradient direction points outwards from the blade, this position is marked as a notch response point. The length of the lateral observation window, the gradient change threshold, and the judgment range from the leading and trailing edges of the edge response operator are also determined experimentally using sample images.

[0062] After processing by the linear response operator and the edge response operator, a set of candidate response points for cracks and scratches and a set of candidate response points for notches are obtained in the blade surface development diagram. For example, in a practical example, when a small chipping occurs in a certain section of the blade leading edge, a continuous high gradient response will appear in the region close to the leading edge in the lateral direction within the corresponding longitudinal coordinate range. The system marks candidate response points for notches at these locations accordingly.

[0063] For pitting and coating defects, a grayscale statistical calculator is used to detect isolated bright or dark spots that are significantly different from the surrounding background within a local area. The grayscale statistical calculator establishes a two-dimensional local statistical window for each pixel in the unfolded image of the blade surface, and calculates the mean grayscale value and grayscale variation range of all pixels within this local statistical window. The grayscale variation range is represented by the difference between the maximum and minimum grayscale values. Simultaneously, several adjacent local statistical windows are selected on the unfolded image as comparison objects. The system compares the mean grayscale value of the current local statistical window with the mean grayscale values ​​of adjacent local statistical windows. If the mean grayscale value of the current local statistical window is significantly higher or lower than the mean grayscale value of adjacent local statistical windows, and there are pixels within the current local statistical window whose grayscale values ​​are close to the upper or lower limit of the allowable range of the industrial camera, then these pixels within the local statistical window are marked as candidate response points for pitting and coating defects.

[0064] The threshold values ​​for significantly higher or lower values ​​in the grayscale statistical calculator are determined through statistical analysis of surface development images of multiple batches of qualified blades and abnormal blades containing defects such as pitting and coating peeling. This ensures that the grayscale statistical calculator will not generate large-scale false alarms under normal texture variations, but will produce a stable response at locations where pitting or coating peeling actually exists. For example, if a coating peeling occurs in the middle region of a blade due to a manufacturing issue, the peeling area will appear as a region with significantly lower brightness than the surrounding area and a more concentrated brightness variation in the blade surface development image. The grayscale statistical calculator can identify this area as a candidate response region for coating defects.

[0065] After obtaining candidate response points for cracks and scratches, candidate response points for notches, and candidate response points for pitting and coating defects, connectivity analysis is performed on these response points in the unfolded map of the blade surface. According to the vertical and horizontal adjacent pixel relationships, response points that are spatially connected or within a preset proximity range are merged into defect candidate connected regions. For connected regions that are long and narrow in the longitudinal direction, the system marks them as candidate regions for cracks and scratches based on the ratio of their longitudinal length to their transverse width. For connected regions that are close to the leading or trailing edge, have a similar transverse width and longitudinal length, and are concentrated near the leading or trailing edge, they are marked as candidate regions for notches. For connected regions that are clustered within a local statistical window, have a limited area, and whose gray-scale mean deviates significantly from the surrounding area, they are marked as candidate regions for pitting or coating defects. For each connected region, the system calculates its circumscribed rectangle boundary, area, longitudinal coordinate range, and transverse coordinate range, and saves this information along with the defect type label as a multi-type defect region. All multi-type defect regions are combined into a multi-type defect region set, providing input for further fine classification and defect parameter extraction using a deep learning classification model.

[0066] Step 6: Combine the geometric parameters obtained from the skeleton and outer contour, as well as the type, size, and position of various defect regions in the quality analysis coordinate system, into a blade quality feature vector, and output the detection results corresponding to the blade quality feature vector. Specifically, this is implemented as follows: After extracting the blade's outer contour and blade skeleton, the chord length, span, cross-sectional width, bending deviation, and shape parameters of the skeleton along its length are calculated based on the completed outer contour and blade skeleton. These geometric parameters are then combined into a geometric parameter sub-vector. Specifically, at each longitudinal sampling position, the image coordinate difference between the leading edge contour line and the trailing edge contour line is taken and converted into a distance in physical space using the industrial camera calibration relationship. This distance is used as the chord length of the cross-section. Along the blade skeleton direction, the physical distance from the blade root to the blade tip is used as the span. On several typical cross-sections, the cross-sectional width distribution is obtained by measuring the distances at multiple lateral positions within the cross-section. Representative cross-sectional width parameters are then extracted according to preset rules. Several typical cross-sections can be selected equidistantly along the blade's length direction. The number of cross-sections is preset to several fixed values ​​based on the blade length and machining accuracy requirements, such as the maximum or average cross-sectional width of each typical cross-section, forming a cross-sectional width parameter group.

[0067] Bending deviation is calculated by comparing the physical coordinates of the blade skeleton with an ideal straight line or a designed skeleton, and determining the lateral offset at each sampling position along the blade length direction. The set of these offsets constitutes the bending deviation sequence.

[0068] The shape parameters of the skeleton along the length direction are obtained by simplifying the description of the bending deviation sequence, such as taking representative offset values ​​at several positions or calculating the average offset of each segment in a segmented manner, to reflect the overall bending trend of the blade.

[0069] According to the preset parameter order, the chord length, span, multiple cross-sectional width parameters, several representative values ​​of bending deviation, and the shape parameters of the skeleton along the length direction are arranged in sequence to form a one-dimensional geometric parameter sub-vector. The meaning of each position in this sub-vector is fixed during system design and does not change with batches, which facilitates the subsequent quality assessment rule base or quality assessment model to read the corresponding geometric parameters according to the position.

[0070] It should be noted that quality assessment can be achieved solely through a rule base, while a quality assessment model is an optional implementation method to improve the consistency and automation of assessment when the sample size is large.

[0071] For a set of multi-type defect regions, the defect type label, defect area, defect length, defect width, defect principal axis direction angle, and longitudinal and lateral coordinates in the quality analysis coordinate system on the blade surface are calculated for each multi-type defect region. These parameters are then combined into a defect parameter sub-vector. The defect type label is determined by the output of the aforementioned deep learning-based classification model, using a discrete category number or category name to represent cracks and scratches, notches, pitting, or coating defects. Deep learning classification models can use convolutional neural networks, which contain several convolutional layers, pooling layers, and fully connected layers. The input is a normalized defect image patch, and the output is the defect category probability and pixel-level defect mask. The training data comes from leaf images with manually labeled defect types and contours. The training process includes dataset partitioning, loss function selection, iterative optimization, etc.

[0072] The defect area is obtained by summing the actual physical areas of all pixels within the defect region on the unfolded image of the blade surface. Specifically, the physical areas corresponding to each defect pixel are added together to obtain the total area of ​​the defect region.

[0073] The defect length and width are obtained by calculating the physical dimensions of the circumscribed rectangle of the defect area in the longitudinal and transverse directions. The defect principal axis direction angle is obtained by calculating the principal inertial direction inside the defect area and finding the angle between this direction and the blade skeleton direction. The angle is expressed as an angle value.

[0074] The longitudinal coordinate of the defect area in the quality analysis coordinate system on the blade surface can be represented by the average longitudinal coordinate of all pixels in the defect area, and the lateral coordinate can be represented by the average lateral coordinate of all pixels in the defect area. Thus, a pair of longitudinal and lateral coordinates can be used to describe the approximate location of the defect on the blade.

[0075] The defect type, defect area, defect length, defect width, defect principal axis angle, longitudinal coordinate, and lateral coordinate are arranged sequentially according to a preset order to form a defect parameter subvector. For the same blade, if there are multiple defect areas of different types, on the one hand, a separate defect parameter subvector can be reserved for each defect; on the other hand, statistics can be summarized by defect type. For example, the total number of defects, the total defect area, the maximum defect length, and the defect location range can be calculated separately for four types: cracks and scratches, notches, pitting, and coating defects. These summarized statistical values ​​are then combined into a defect summary subvector. Each statistical item in the defect summary subvector has a clear meaning and a fixed arrangement order, used to comprehensively characterize the overall quality of the blade under different defect types. For example, in a simple real-world scenario, if there are three small notches and several pits in the middle section of the leading edge of a blade, the defect summary subvector will record that the number of notches is three, the total area of ​​notches is a small value, the notch location is concentrated in the middle section of the longitudinal coordinate, and the lateral coordinate is close to the leading edge, while the number and area of ​​pitting types are counted separately.

[0076] After constructing the geometric parameter subvector and the defect summary subvector, the geometric parameter subvector and the defect summary subvector are concatenated in a preset order to form the blade quality feature vector. The blade quality feature vector is a one-dimensional parameter sequence. The first part of the parameters corresponds to the geometric parameters of the blade, and the second part of the parameters corresponds to the statistical characteristics of multiple types of defects, which are used as input to the quality assessment rule base or quality assessment model.

[0077] The quality assessment rule base can be pre-defined by process engineers and quality engineers. The rule base lists the allowable ranges for geometric parameters such as bending deviation, cross-sectional width, and span, as well as the allowable ranges for the number and area of ​​different defect types under different quality levels. For example, it stipulates that a high-level defect is defined as one where the bending deviation is within the allowable range and the total area of ​​the defects is below a certain empirical threshold; an intermediate-level defect is defined as one where the bending deviation is close to the upper limit or the total area of ​​the defects is close to the empirical threshold; and a low-level defect or a defect that is unacceptable is defined as one where the bending deviation exceeds the allowable range or the length of a single defect exceeds the empirical limit. The quality assessment model can be trained using machine learning methods. For example, it can use the blade quality feature vectors of historical blades and the corresponding manual quality assessment results as samples to obtain the mapping relationship from the blade quality feature vectors to the quality level and treatment recommendations.

[0078] During operation, the blade quality feature vector of each blade is input into the quality assessment rule base and quality assessment model to obtain the blade's quality grade and corresponding handling recommendations. Handling recommendations can include options such as allowing direct assembly and use, requiring grinding and repair followed by re-inspection, or requiring scrapping. The system simultaneously organizes blade identification information, blade quality feature vector, quality grade, and handling recommendations into structured records and writes them into the quality database. The inspection terminal displays the inspection results, including the quality grade, geometric parameter summary, and defect list. The geometric parameter summary can select key geometric parameters such as the maximum bending deviation, typical cross-sectional width, and span. The defect list can enumerate the defect type, size, and location of various defect areas, allowing operators to quickly understand the blade's quality condition.

[0079] For example, in a certain actual production batch, if the number of pits in the middle section of the leading edge of the same type of blade repeatedly approaches the upper limit in multiple batches, engineers can use the blade quality feature vector and defect list recorded in the quality database to perform statistical analysis, trace the corresponding process, and optimize the coating or curing process.

[0080] It should be noted that the above thresholds were selected by statistically analyzing multiple batches of samples and combining the false detection rate and false negative rate under different quality levels. These are implementation parameters that can be adjusted according to equipment and process conditions, such as empirical thresholds and preset ranges.

[0081] This invention uses a preview imaging statistics and imaging parameter configuration table to adaptively adjust the exposure time, industrial camera gain, and controllable light source combination. It also combines extreme area masking, multi-frame compensation, and local brightness equalization to reduce the impact of strong reflections and shadows on the image quality of the blades, and stably obtain contour enhancement maps and texture enhancement maps under complex lighting conditions. Based on the blade surface development diagram and quality analysis coordinate system constructed from the model design outline and refined skeleton, geometric measurement and defect location of three-dimensional curved blades are realized in a unified two-dimensional coordinate system. This makes the results of different batches and different inspection stations comparable and traceable. Then, by extracting multiple types of candidate regions for cracks, notches, pits and coating defects on the development diagram, and organizing the geometric parameters and defect parameters into a blade quality feature vector, the system outputs quality level and treatment suggestions in conjunction with a quality assessment rule base or quality assessment model. This achieves unified quantitative evaluation of multiple types of defects and geometric deviations, improves the defect detection rate and judgment consistency, reduces manual dependence, and is suitable for large-scale online quality inspection of ventilation equipment blades.

[0082] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0083] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0084] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0086] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting the quality of ventilation equipment blades based on image processing, characterized in that: The specific steps include: At the inspection station, an industrial camera and a controllable light source are used to acquire images of the blades. The exposure and light source are adjusted according to the brightness and saturation areas of the image to obtain the original image of the blades. The original leaf image is preprocessed for illumination and reflection, highlight and shadow areas are marked, local brightness equalization is applied to the remaining areas, and multi-frame information is used to compensate for the highlight and shadow areas to generate contour enhancement map and texture enhancement map. Using the contour enhancement map as input, the blade region is located and the outer contour is extracted. The broken contour is aligned and completed by combining the model design contour. The skeleton along the length direction is extracted within the outer contour. Using the skeleton as the vertical reference, the blade region in the texture enhancement map is resampled along the skeleton and normal directions to construct the blade surface unfolding map and establish a quality analysis coordinate system with the skeleton position and the distance to the leading and trailing edges as coordinates. In the unfolded image of the blade surface, response operators along the skeleton and normal, as well as gray-scale operators, are used to extract candidate regions for cracks, notches, pits, and coating defects, resulting in multiple types of defect regions. The geometric parameters obtained from the skeleton and outer contour, as well as the type, size, and position of various defect regions in the quality analysis coordinate system, are combined to form the blade quality feature vector, and the detection results corresponding to the blade quality feature vector are output.

2. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 1, characterized in that: The steps involved in acquiring blade images using an industrial camera and a controllable light source at the inspection station, adjusting the exposure and light source based on the image brightness and saturation areas, and obtaining the original blade image include: The brightness histogram, high-brightness saturated pixel ratio, and low-brightness pixel ratio are calculated from the preview image captured when the blade enters the inspection station. The obtained statistics are matched with the preset imaging parameter configuration table. The corresponding exposure time, industrial camera gain, and controllable light source combination are selected from the imaging parameter configuration table. The parameters of the industrial camera and controllable light source are set. After the parameter settings are completed, the original image of the blade is captured.

3. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 2, characterized in that: The steps for preprocessing the original leaf image by considering illumination and reflection, marking highlight and shadow areas, and performing local brightness equalization on the remaining areas to generate contour enhancement and texture enhancement maps include: Highlight and shadow areas are detected based on pixel brightness threshold and local gradient threshold in the original image of the leaf, and extreme region masks are generated according to the connected regions. Pixel areas outside the extreme region mask are regarded as non-extreme regions. Local brightness equalization and local contrast stretching are performed on non-extreme regions. The images of non-extreme regions after local brightness equalization and local contrast stretching are used as the base images for contour enhancement and texture enhancement.

4. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 3, characterized in that: The steps for compensating for highlights and shadows using multi-frame information include: While keeping the position of the industrial camera unchanged, multiple frames of original images of the blades are acquired by sequentially switching between different polarization directions or different combinations of light sources. Spatial registration is performed on multiple frames of original leaf images. In the highlight areas marked by the mask in extreme regions, the pixels with the lowest saturation or brightness within a preset range are selected to replace the corresponding pixels. Interpolation compensation is performed within the shadow area marked by the mask in the extreme region using pixels from adjacent non-extreme regions and texture pixels at the corresponding positions of the previous batch of qualified blades to obtain the compensated original image of the blade used to generate the contour enhancement map and texture enhancement map.

5. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 1, characterized in that: The steps of taking the contour enhancement map as input, locating the blade region and extracting the outer contour, aligning and completing the broken contour in conjunction with the model design contour, and extracting the skeleton along the length direction within the outer contour include: In the contour enhancement map, the initial position and size of the blade region are obtained through target detection or template matching. In the blade region, edge detection and threshold segmentation are used to extract the initial outer contour. The initial outer contour is aligned with the design contour corresponding to the blade model and the scale is normalized. At the location where there is a break in the initial outer contour, the contour fragments at the corresponding positions of the design contour are used to complete and smooth it. In the interior of the completed outer contour, the blade skeleton distributed along the blade length direction is extracted based on distance transformation and thinning algorithms.

6. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 1, characterized in that: The steps include: using the skeleton as the vertical reference, resampling the blade region in the texture enhancement map along the skeleton and normal directions, constructing the blade surface unfolding map, and establishing a quality analysis coordinate system with the skeleton position and the distance to the leading and trailing edges as coordinates. In the texture enhancement map, the blade region is divided into several sampling positions along the skeleton direction using the blade skeleton. At each sampling position, the normal direction sampling line from the leading edge to the trailing edge is determined according to the outer contour. The pixel arrays arranged in the longitudinal and lateral positions are obtained by resampling along the sampling line at a preset interval. The pixel arrays are used to form a unfolded map of the blade surface. The quality analysis coordinate system of the blade surface is established by using the sampling index along the blade skeleton as the longitudinal coordinate and the normalized distance from the blade skeleton to the leading edge and from the blade skeleton to the trailing edge as the lateral coordinate.

7. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 1, characterized in that: The steps for extracting candidate regions for cracks, notches, pits, and coating defects in the blade surface development map using response operators along the skeleton and normal, as well as grayscale calculation operators, to obtain multi-type defect regions include: In the unfolded image of the blade surface, a linear response operator is applied along the longitudinal direction to extract slender strip-shaped high-response regions as candidate regions for cracks and scratches. An edge response operator is applied along the transverse direction to extract edge morphology change regions near the leading and trailing edges as candidate regions for notches. The gray-level mean and gray-level variance are calculated within a local window and compared with adjacent windows. Isolated regions with gray-level anomalies are extracted as candidate regions for pitting and coating defects. The candidate regions are clustered according to connected domains and the region boundaries are recorded to form a set of multi-type defect regions.

8. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 7, characterized in that: After obtaining the multi-type defect regions, the steps for classifying and extracting parameters from these regions include: Based on the boundary of each multi-type defect region, defect image blocks are cropped from the unfolded image of the blade surface. Size normalization and grayscale normalization are performed on the defect image blocks. The normalized defect image blocks are then input into a pre-trained deep learning classification model. The deep learning classification model outputs defect type labels and defect contour masks. Based on the defect contour masks, the area, length, width, and principal axis angle of the defect region are calculated. The defect type labels, area, length, width, and principal axis angle are recorded in a multi-type defect region set.

9. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 8, characterized in that: The steps for combining the geometric parameters obtained from the skeleton and outer contour, as well as the type, size, and position of various defect regions in the quality analysis coordinate system, into a blade quality feature vector include: The chord length, span, cross-sectional width, bending deviation, and shape parameters of the skeleton along the length direction are combined into a geometric parameter subvector; The defect type label, defect area, defect length, defect width, defect principal axis direction angle, and longitudinal and transverse coordinates in the quality analysis coordinate system of each multi-type defect region are combined to form a defect parameter sub-vector. All defect parameter subvectors of the same blade are concatenated or compressed in a preset order to obtain a defect summary subvector. The geometric parameter subvectors are then concatenated with the defect summary subvector to form the blade quality feature vector.

10. The method for detecting the quality of ventilation equipment blades based on image processing according to claim 9, characterized in that: The detection results corresponding to the output blade quality feature vector include: Input the blade quality feature vector into the quality assessment rule base or quality assessment model, and determine the blade quality grade and corresponding treatment recommendations based on the value range of bending deviation, cross-sectional width and span in the geometric parameter sub-vector and the combination relationship of defect type, defect size and defect location distribution in the defect summary sub-vector. The blade identification information, blade quality feature vector, quality grade, and treatment recommendations are organized into structured records and written into the quality database. The test results, which include quality grade, geometric parameter summary, and defect list, are output on the test terminal.

Images