Intelligent detection method and system for surface defects of automobile air conditioner compressor piston based on visual analysis
By acquiring images from multiple perspectives and identifying defects using deep learning models, the problem of traditional methods being unable to cover the entire circumferential surface of the piston was solved, achieving efficient and accurate defect detection and generating detailed three-dimensional defect reports.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU JIANGDONG INTERNAL COMBUSTION ENGINE FITTINGS
- Filing Date
- 2026-02-27
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies struggle to achieve complete coverage imaging of the entire circumferential surface of an automotive air conditioning compressor piston, especially in areas with high curvature. Furthermore, traditional methods are prone to introducing positioning errors and efficiency losses.
At least three sets of ring camera arrays are used for multi-view image acquisition. The macro camera and industrial camera are arranged at different pitch angles. The structured light 3D scanning and multi-focus depth-of-field synthesis are combined. Defects are identified by multi-scale feature fusion deep learning model. The 3D coordinates are generated by camera calibration and triangulation.
It achieves complete imaging of the piston's full circumferential surface, overcomes the blind zone in high curvature areas, reduces positioning errors and false detection rates, improves the detection rate and generalization ability of minute defects, and provides an intuitive view of the spatial distribution of defects and data support.
Smart Images
Figure CN122115391A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of product quality inspection technology, specifically to an intelligent detection method and system for surface defects on the piston of an automotive air conditioning compressor based on visual analysis. Background Technology
[0002] With the rapid development of the automotive industry and the increasing demands of consumers for vehicle performance, comfort, and reliability, the quality control of core components in automotive air conditioning systems, as one of the key subsystems, is receiving increasing attention. The air conditioning compressor, as the power source of the refrigeration cycle, has internal piston components that bear significant mechanical stress and thermal load during long-term, high-frequency reciprocating motion. Its surface quality (such as the presence of scratches, dents, cracks, burrs, or oxidation defects) directly affects the compressor's sealing performance, operating efficiency, and even its overall lifespan.
[0003] Currently, the industry still mainly relies on manual visual inspection or traditional machine vision methods based on simple image thresholding for the detection of surface defects in compressor pistons. Manual inspection methods suffer from problems such as strong subjectivity, low efficiency, and susceptibility to fatigue and missed detections, making it difficult to meet the requirements of modern intelligent manufacturing for high consistency and high-speed production. While traditional machine vision methods can achieve automation to a certain extent, they are sensitive to interference factors such as changes in lighting, surface reflections, and complex textured backgrounds, and lack robust recognition capabilities for small, irregular, or low-contrast defects, resulting in poor generalization performance and high false positive and false negative rates.
[0004] Chinese invention patent CN115290662A discloses a multi-view visual surface defect detection system and its detection method. The detection system includes: a data acquisition end for acquiring images of parts and uploading the image data to an industrial control terminal; and an industrial control terminal for detecting the images acquired by the acquisition end. The acquisition end consists of multiple cameras arranged in a spatial array to form a multi-view system, with multiple cameras simultaneously acquiring images of the parts. The array can be arranged in a straight line or an arc, and its spatial position can be arbitrarily determined according to specific circumstances. The industrial control terminal uses video stream analysis to perform real-time analysis of the time-series images of the parts simultaneously acquired by the multiple cameras, forming a two-dimensional relationship diagram with the positions of the multiple cameras, and accurately determining the location of defects for defect localization and saving relevant defect detection image data.
[0005] However, the above and similar technical solutions still have the following shortcomings: Since the piston of the flat displacement air conditioning compressor of automobile is a slender rod structure with ball heads or conical surfaces at both ends, it is difficult to achieve blind-angle imaging by using a fixed viewing angle imaging method when completely covering its entire circumferential surface (especially the axial blind area and the high curvature transition area such as the concave chamfer). Although mechanical flipping or rotating fixtures can improve coverage, they will reduce the detection cycle and introduce positioning errors. Summary of the Invention
[0006] The purpose of this invention is to provide a method and system for intelligent detection of surface defects on the piston of an automotive air conditioning compressor based on visual analysis, so as to solve the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent detection method for surface defects of automotive air conditioning compressor pistons based on visual analysis, comprising: S1: Multi-view image processing: Using at least three sets of ring camera arrays, full-surface images of the piston's upper ball head area, middle rod body area and lower conical surface area are acquired to obtain the corresponding full-surface images of the piston; S2: Intelligent Defect Recognition: The full surface image of the piston is used as the input of a multi-scale feature fusion deep learning model, and the corresponding defect is obtained as the output. S3: Defect Mapping Determination: Based on the corresponding camera coefficients in the ring camera array, the identified defects are mapped to the three-dimensional digital model of the piston to obtain the corresponding defect detection report.
[0008] Furthermore, the corresponding full surface image of the piston is obtained, including: S1.1: Camera distribution: Macro cameras are set at the entrance and exit of the closed imaging box through camera mounting rings, and multiple industrial cameras are set in the middle of the closed imaging box, and the pitch angle of the macro cameras is greater than that of the industrial cameras. S1.2: Image Acquisition: An illumination unit is set on the camera mounting ring, and through information interaction with the central controller, under the illumination of the illumination unit, the macro camera acquires the full surface image of the upper ball head area and the full surface image of the lower cone area of the piston, and the industrial camera acquires the full surface image of the middle rod area of the piston. S1.3: Image Enhancement: By using a structured light 3D scanning algorithm and multi-focus depth-of-field synthesis, image enhancement processing is performed on the full surface images of the upper ball head region and the lower cone region of the piston to obtain a full surface image with enhanced details.
[0009] Furthermore, the pitch angle of the macro camera is greater than 70°, and the pitch angle of the industrial camera is set to 10°-30°.
[0010] Furthermore, the illumination unit corresponding to the macro camera is equipped with a ring light or a strip light, and the illumination unit corresponding to the industrial camera is equipped with a dome light or a coaxial light, and the ring light or strip light is set on the same side or opposite side of the macro camera.
[0011] Furthermore, a full-surface image with enhanced details is obtained, including: S1.3.1: Multi-focus fusion: The shooting position of the macro camera is adjusted by a piezoelectric ceramic actuator, and full-surface images corresponding to different positions are acquired according to the adjusted shooting position. At the same time, the full-surface images corresponding to different positions are aligned at the pixel level by an image registration algorithm to obtain the sharpness score of each pixel. The pixels corresponding to the maximum sharpness score are stitched together to obtain the corresponding fused full-surface image. S1.3.2: Structured light reconstruction: During the macro camera shooting process, a light spot pattern is projected, a full surface image with deformed light spots is acquired, and the gray value corresponding to each pixel in the full surface image with deformed light spots is determined. At the same time, based on the gray value of each pixel in different full surface images with deformed light spots, the corresponding phase value is determined, the corresponding three-dimensional spatial coordinates are obtained, and the corresponding three-dimensional point cloud map is constructed based on the three-dimensional spatial coordinates. S1.3.3: Image fusion: The fused full-surface image and the three-dimensional point cloud image are fused at the pixel level to obtain a full-surface image with enhanced details.
[0012] Furthermore, the output retrieves the corresponding identified defects, including: S2.1: Model Setup: A multi-scale feature fusion deep learning model is constructed by setting up an improved deep learning network with sequentially connected dual input terminals, backbone network, neck network, and detection head. S2.2: Model Prediction: Using dual input terminals, the full surface image of the piston's middle rod region and the full surface image with enhanced details are used as inputs to a multi-scale feature fusion deep learning model, and the corresponding defect identification results are obtained as outputs.
[0013] Furthermore, the backbone network is provided with a high-resolution feature branch, and the neck network is provided with a high-resolution thinning branch. Both the high-resolution feature branch and the high-resolution thinning branch are used to process the full-surface image after detail enhancement.
[0014] Furthermore, the corresponding defect detection report is obtained, including: S3.1: Coordinate Determination: All cameras in the circular camera array are calibrated to obtain the intrinsic and extrinsic parameters corresponding to each camera. Based on the intrinsic and extrinsic parameters corresponding to the cameras, the preliminary three-dimensional coordinates of each defect information under different cameras are determined. At the same time, the preliminary three-dimensional coordinates of each defect information are combined by triangulation to determine the final three-dimensional coordinates corresponding to each defect information. S3.2: Data Deduplication: Based on the final three-dimensional coordinates corresponding to each defect information, the Euclidean distance between two adjacent defect information is obtained. Based on the comparison between the Euclidean distance and the preset fusion threshold, the defect identification results are deduplicated. The deduplicated defect identification results are then mapped in the piston three-dimensional digital model to obtain the corresponding defect detection report with three-dimensional coordinates.
[0015] Furthermore, based on the comparison between the Euclidean distance and the preset fusion threshold, the defect identification results are deduplicated, specifically as follows: When the Euclidean distance is greater than the preset fusion threshold, the corresponding defect information is not fused, and the final three-dimensional coordinates corresponding to the defect information are the predicted three-dimensional coordinates; otherwise, the two corresponding defect information are fused, and one defect information is retained among the two adjacent defect information, and the final three-dimensional coordinates corresponding to the retained defect information are the predicted three-dimensional coordinates.
[0016] The intelligent detection system for surface defects of automotive air conditioning compressor pistons based on visual analysis uses any one of the aforementioned intelligent detection methods for surface defects of automotive air conditioning compressor pistons based on visual analysis.
[0017] Compared with the prior art, the beneficial effects of the present invention are: Firstly, this invention uses at least three sets of ring camera arrays to acquire multi-view images of the upper ball head area, the middle rod body area, and the lower conical surface area of the piston. This solves the problem of blind spots in traditional fixed-view imaging, which is difficult to cover high curvature areas. At the same time, the macro camera is arranged at a high pitch angle and the industrial camera is arranged at a low pitch angle to cover the circumferential surface of the rod body. This ensures complete imaging of the entire circumferential surface of the slender rod piston, avoiding positioning errors and efficiency losses introduced by mechanical flipping or rotating fixtures. Secondly, this invention uses structured light 3D scanning and multi-focal depth-of-field fusion algorithms to perform pixel-level fusion of multi-focal images obtained from high curvature areas such as spherical heads and conical surfaces, and reconstructs and generates 3D point cloud maps, thereby synthesizing RGB-D images with enhanced details. This not only effectively overcomes surface reflection, texture interference and depth-of-field limitations, but also improves the detection rate of small, low-contrast defects. Thirdly, this invention uses a dual-input deep learning model to process the pole area image and the high-resolution image after detail enhancement, and performs feature interaction and fusion through the high-resolution feature branch in the backbone network and the thinning branch in the neck network, thereby achieving collaborative recognition of defects of different scales, effectively improving the generalization ability of complex defects, and reducing the false detection and false detection rates. Fourthly, this invention calculates the coordinates of defects in three-dimensional space through camera calibration and triangulation, and performs deduplication and fusion of defects repeatedly detected from multiple perspectives based on the Euclidean distance threshold. At the same time, the defect information is mapped to the three-dimensional digital model of the piston to generate an inspection report containing defect type, coordinates, size and confidence level. This not only provides an intuitive view of the spatial distribution of defects, but also provides data support for subsequent process adjustments and quality traceability. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the intelligent detection method for surface defects of automotive air conditioning compressor pistons in this invention. Figure 2 This is a flowchart illustrating the multi-view image processing method of the present invention; Figure 3 This is a schematic diagram illustrating the process of acquiring the full-surface image with enhanced details in this invention; Figure 4 This is a flowchart illustrating the intelligent defect identification method of the present invention; Figure 5 This is a flowchart illustrating the defect mapping determination method in this invention. Detailed Implementation
[0019] 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.
[0020] refer to Figure 1 This embodiment provides a visual analysis-based intelligent detection method for surface defects on the piston of an automotive air conditioning compressor. The method includes the following steps: Step S1: Multi-view image processing. At least three sets of circular camera arrays are sequentially set up according to the axis of the piston's conveying direction to cover the upper ball head area, the middle shaft area, and the lower conical surface area of the piston, respectively, thereby acquiring the corresponding full-surface images of the piston. Simultaneously, structured light 3D scanning and multi-focus depth-of-field synthesis algorithms are used to reconstruct and synthesize the full-surface images corresponding to the upper ball head area and the lower conical surface area of the piston, in order to reconstruct and enhance the morphology of high-curvature areas, obtaining full-surface images with enhanced details.
[0021] Specifically, the annular camera array covering the upper ball head region and the lower conical surface region of the piston is equipped with macro cameras arranged at high elevation angles to obtain full surface images of the corresponding upper ball head and its root chamfered area, as well as the corresponding full surface images of the lower conical surface and the concave transition area. In this embodiment, the high elevation angle corresponds to an angle greater than 70°. Simultaneously, the annular camera array located in the middle rod region of the piston is equipped with macro cameras arranged at low elevation angles to obtain full surface images of the corresponding circumferential surface of the middle rod region of the piston. The angle of the annular camera array located in the middle rod region of the piston is smaller than the angle of the annular camera array located in the upper ball head region and the lower conical surface region of the piston.
[0022] Step S2: Intelligent Defect Recognition. The enhanced full-surface images (including the enhanced full-surface images of the upper ball head region and the lower cone region of the piston) obtained in Step S1 are combined with the full-surface image of the middle rod region of the piston and used as input to the multi-scale feature fusion deep learning model to output the corresponding identified defects.
[0023] Furthermore, the multi-scale feature fusion deep learning model in this embodiment is configured with a multi-branch feature extraction structure. This multi-branch feature extraction structure includes at least one branch for processing the full-surface image features after detail enhancement, and the branch for processing the full-surface image features after detail enhancement interacts and fuses with the branch for processing the full-surface image features of the piston's central shaft region.
[0024] Step S3: Defect Mapping Determination. Based on the calibrated camera array coefficients from Step S1, the defect information identified in Step S2 is mapped onto the piston's 3D digital model. Furthermore, duplicate defects identified from different viewpoints are deduplicated and fused to obtain a corresponding defect detection report with 3D coordinates.
[0025] This embodiment also provides a visual analysis-based intelligent detection system for surface defects on the piston of an automotive air conditioning compressor, which uses the aforementioned visual analysis-based intelligent detection method for surface defects on the piston of an automotive air conditioning compressor.
[0026] In this embodiment, at least three sets of ring camera arrays are used to acquire full-surface images corresponding to the upper ball head region, the middle shaft region, and the lower conical region of the piston. The full-surface images corresponding to the upper ball head region and the lower conical region are then reconstructed and enhanced. (Reference) Figure 2 This embodiment provides a multi-view image processing method, which specifically includes the following steps: Step S1.1: Camera Distribution. The piston is fixed and guided on a linear guide rail or conveyor belt using V-blocks or custom fixtures (which can be customized according to actual needs, and are not specifically described in this embodiment) to ensure that the piston's axis is consistent with the conveying direction. Simultaneously, a closed imaging box is set on the linear guide rail or conveyor belt to acquire images of the piston's full surface using a set camera array within the closed imaging box.
[0027] Furthermore, the enclosed imaging chamber contains at least three camera mounting rings to capture images of the entire piston surface as it passes through the chamber. Specifically, in this embodiment, three camera mounting rings are provided: a first ring at the entrance of the enclosed imaging chamber, a second ring in the middle, and a third ring at the exit. Both the first and third rings are secured with macro cameras via universal adjustment brackets (including locking knobs with multiple degrees of freedom such as left / right, up / down, pitch, and yaw). The macro cameras on the first and third rings are symmetrically distributed to capture images of the entire surface of the upper ball head region and the lower conical region of the piston. Additionally, a second ring contains multiple industrial cameras (e.g., 4-8) evenly distributed to cover the entire circumferential surface of the piston's central shaft. In other words, multiple industrial cameras evenly distributed on the second camera mounting ring are used to acquire full-surface images of the piston's central rod region.
[0028] Furthermore, in this embodiment, the pitch angles of the macro cameras mounted on the first and third camera mounting rings are both greater than 70°. This means the optical axis of the macro camera lens is greater than 70°, and the macro camera is tangentially aligned with the upper ball head root region and the lower conical surface and concave transition area of the piston. It is worth noting that the tangential alignment of the macro cameras in this embodiment is an approximate alignment, not a true physical alignment. Meanwhile, in this embodiment, the pitch angle of the industrial camera mounted on the second camera mounting ring is set to 10°-30°. This means the angle between the optical axis of the industrial camera lens and the horizontal plane is set to 10°-30°, to cover the entire circumferential surface of the piston's middle rod from different angles.
[0029] Step S1.2: Image Acquisition. A photoelectric encoder or through-beam laser photoelectric sensor is installed on the outside of the enclosed imaging chamber to determine the corresponding positional relationship between the piston on the linear guide or conveyor belt and the enclosed imaging chamber. Based on the determined positional relationship, the camera array set in step S1.1 acquires the corresponding full-surface image of the piston. Specifically, the photoelectric encoder or through-beam laser photoelectric sensor interacts with the built-in controller of the camera array through the central controller. The information exchange between the photoelectric encoder / through-beam laser photoelectric sensor and the central controller, and between the central controller and the built-in controller of the camera array, is all conducted via wireless communication (e.g., 5G). In other words, when the piston on the linear guide or conveyor belt is transported to the outside of the enclosed imaging chamber, the photoelectric encoder or through-beam laser photoelectric sensor acquires the corresponding positional information and transmits it to the central controller. Simultaneously, the central controller sends a running command to the built-in controller of the camera array based on the positional information, controlling each camera in the camera array to start and acquire the corresponding full-surface image of the piston.
[0030] Furthermore, in this embodiment, each camera mounting ring is equipped with a corresponding illumination unit. The first and third camera mounting rings each have ring lights or strip lights positioned on the same side or opposite side of the macro camera to create a noticeable shadow effect on the tiny scratches and pits at the base of the piston head and the concave area, enhancing contrast. Simultaneously, the second camera mounting ring has dome lights or coaxial lights to eliminate highlights caused by reflections from the cylindrical surface of the piston rod. In other words, through the ring lights or strip lights, the macro cameras mounted on the first and third camera mounting rings acquire full-surface images of the upper ball head area and the lower conical surface area of the piston. Through the dome lights or coaxial lights, the industrial camera mounted on the second camera mounting ring acquires full-surface images of the middle piston rod area.
[0031] Step S1.3: Image Enhancement. This involves using a structured light 3D scanning algorithm and multi-focus depth-of-field synthesis to enhance the full-surface images of the upper ball head region and the lower conical region of the piston acquired in Step S1.2, obtaining corresponding full-surface images with enhanced details. (Reference) Figure 3 The details are as follows: Step S1.3.1: Multi-focus fusion. This involves placing a piezoelectric ceramic actuator on a gimbal with a fixed macro camera, and this actuator can move along the Z-axis (i.e., the optical axis of the camera lens). Simultaneously, a piezoelectric ceramic controller inside the actuator interacts with the central controller. Specifically, while the central controller sends operating commands to the built-in controller of the camera array, it also sends corresponding operating commands to the piezoelectric ceramic controller. In other words, the piezoelectric ceramic controller adjusts the position of the macro camera based on the received operating commands, and then uses the adjusted macro camera to acquire the corresponding full-surface image of the piston.
[0032] Furthermore, based on the preset adjustment frequency corresponding to the piezoelectric ceramic actuator (which can be specifically set according to actual needs, so it is not specifically described in this embodiment), the full surface images of the piston at different positions on the universal adjustment bracket are acquired, that is, the full surface images of the upper ball head area and the lower cone area of the piston at multiple different positions are acquired.
[0033] Furthermore, an image registration algorithm is used to perform pixel-level alignment on the full-surface images of the upper ball head region (and the lower cone region) of the piston at different locations. Specifically, each acquired full-surface image is divided according to a set pixel size, and each full-surface image is pixel-level aligned based on the pixels within each divided image. Simultaneously, a high-pass filter (such as a Laplacian operator or gradient energy) is used to determine the sharpness score corresponding to each pixel in each full-surface image (note that in this embodiment, the sharpness score can be determined by filtering high-frequency information using a high-pass filter to determine the corresponding pixel value). In other words, based on the sharpness scores of each pixel in different full-surface images, the maximum sharpness score corresponding to the same pixel in different full-surface images is determined. Then, each pixel corresponding to the maximum sharpness score is stitched together to synthesize the corresponding fused full-surface image, namely, the fused full-surface image of the upper ball head region of the piston and the fused full-surface image of the lower cone region.
[0034] Step S1.3.2: Structured Light Reconstruction. A miniature digital light processor projection module or laser line source generator is installed on the side of each macro camera. When the macro camera acquires a full-surface image, a preset light spot pattern is projected onto the root of the piston's ball head or the concave transition area through the miniature digital light processor projection module or laser line source generator. It is worth noting that the preset light spot pattern in this embodiment includes, but is not limited to, sinusoidal stripe light. Simultaneously, multiple full-surface images with different deformed light spots are obtained based on the instantaneous state of the piston surface.
[0035] Furthermore, based on multiple full-surface images with different deformed light spots corresponding to the same instantaneous state, each acquired full-surface image with deformed light spots is divided according to a set pixel size. Simultaneously, based on the pixels divided in each full-surface image with deformed light spots, the grayscale value corresponding to each pixel in the full-surface image with deformed light spots is obtained, and the grayscale value of each pixel in different full-surface images with deformed light spots is determined. Specifically, based on the different grayscale values corresponding to each pixel and a phase-shifting algorithm, the phase value corresponding to each pixel is determined. It is worth noting that in this embodiment, the grayscale value of the pixel is processed and the corresponding phase value is determined using a phase-shifting algorithm, which are conventional technical processing methods, and therefore are not specifically described in this embodiment.
[0036] Furthermore, using the geometric model corresponding to the macro camera and the miniature digital light processor projection module (or laser line source generator), each determined phase value is transformed to obtain its corresponding three-dimensional spatial coordinates. (It is worth noting that in this embodiment, determining the three-dimensional spatial coordinates corresponding to the phase value through the geometric model is a conventional technical processing method, and therefore is not specifically described in this embodiment.) Simultaneously, the three-dimensional spatial coordinates obtained for each phase value are combined to construct the corresponding three-dimensional point cloud map.
[0037] Step S1.3.3: Image Fusion. This involves pixel-level fusion of the fused full-surface image obtained in step S1.3.1 and the 3D point cloud obtained in step S1.3.2 to obtain the corresponding enhanced full-surface image. Specifically, using a preset multi-channel image (i.e., a four-channel image), the red, green, and blue components of the preset multi-channel image are filled based on the fused full-surface image obtained in step S1.3.1. Simultaneously, the grayscale values (i.e., height information) of the preset multi-channel image are filled based on the 3D point cloud obtained in step S1.3.2. In other words, based on the filled preset multi-channel image, the corresponding RGB-D image, i.e., the enhanced full-surface image, is obtained.
[0038] In this embodiment, the corresponding identification defects are obtained through a multi-scale feature fusion deep learning model. (Reference) Figure 4 This embodiment provides a defect intelligent identification method, which specifically includes the following steps: Step S2.1: Model Setup. This involves setting up a corresponding multi-scale feature fusion deep learning model based on an improved deep learning network (e.g., improved YOLOv5). In this embodiment, the multi-scale feature fusion deep learning model includes a backbone network, a neck network, and a detection head. The backbone network is connected to the dual input terminals and the neck network, and the neck network is connected to the detection head. Specifically, the dual input terminals input the full-surface image of the piston's central rod region obtained in step S1.2 and the full-surface image with enhanced details obtained in step S1.3.3 into the multi-scale feature fusion deep learning model. The backbone network then performs preliminary feature extraction on both the full-surface image and the enhanced full-surface image. Simultaneously, the neck network performs interactive fusion of the extracted preliminary features, and the detection head identifies the fused features to determine the corresponding recognition result.
[0039] Furthermore, in this embodiment, the backbone network is provided with a high-resolution feature branch, and the neck network is provided with a high-resolution thinning branch. Both the high-resolution feature branch and the high-resolution thinning branch are used to perform feature recognition processing on the full surface image after detail enhancement.
[0040] Step S2.2: Model Prediction. This involves using the dual inputs of the multi-scale feature fusion deep learning model. The full-surface image of the piston's central rod region obtained in Step S1.2 and the enhanced full-surface image obtained in Step S1.3.3 are both used as inputs. The enhanced full-surface image undergoes preliminary feature recognition through a high-resolution feature branch in the backbone network, and the identified features are then transferred to a high-resolution thinning branch in the neck network. Simultaneously, the neck network fuses the features corresponding to the full-surface image of the piston's central rod region and the enhanced full-surface image. The detection head of the multi-scale feature fusion deep learning model analyzes the fused features obtained from the neck network to obtain the corresponding defect identification result.
[0041] In this embodiment, a corresponding defect detection report with three-dimensional coordinates is obtained using the three-dimensional digital model of the piston and the defect identification result obtained in step S2.2. (Reference) Figure 5 This embodiment provides a defect mapping determination method, which specifically includes the following steps: Step S3.1: Coordinate Determination. This involves calibrating all cameras in the circular camera array set in step S1.1 using a 3D calibration board of known dimensions (e.g., a 3D target with coded points) to calculate the intrinsic and extrinsic parameters of each camera. It is worth noting that the camera calibration in this embodiment uses conventional techniques, and therefore, they are not specifically described in this embodiment.
[0042] Furthermore, in the constructed 3D digital model of the piston, the preliminary 3D coordinates of each defect identified in step S2 under different cameras are determined by using the intrinsic and extrinsic parameters corresponding to each camera. Simultaneously, the preliminary 3D coordinates of each defect under different cameras are combined using triangulation to determine the final 3D coordinates of each defect.
[0043] Step S3.2: Data Deduplication. This involves obtaining the Euclidean distance between adjacent defect pieces of information based on the final three-dimensional coordinates of each defect identified in Step S3.1. Simultaneously, the obtained Euclidean distance is compared with a preset fusion threshold (which can be set according to actual needs, but is not specifically described in this embodiment, e.g., 1mm). Based on the comparison result, the defect information identified in Step S2 is deduplicated. Specifically: When the obtained Euclidean distance is greater than the preset fusion threshold, the two corresponding defect information pieces are not fused, and the final 3D coordinates of the defect information are the predicted 3D coordinates. Conversely, when the obtained Euclidean distance is not greater than the preset fusion threshold, the two corresponding defect information pieces are fused. That is, only one defect information piece is retained among two adjacent defect information pieces, and the remaining defect information pieces are deleted. The final 3D coordinates of the retained defect information are then the predicted 3D coordinates.
[0044] Furthermore, based on the predicted 3D coordinates corresponding to the determined defect information, a one-to-one mapping is performed on the 3D digital model of the piston. Simultaneously, based on the mapped 3D digital model, a corresponding defect detection report with 3D coordinates is obtained. Specifically, the defect detection report in this example includes the defect ID, defect type, defect 3D coordinates, defect size, defect confidence level, the region where the defect is located, and the associated image of the defect.
[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.
Claims
1. A visual analysis-based intelligent detection method for surface defects on the piston of an automotive air conditioning compressor, characterized in that, Including: S1: Multi-view image processing: Using at least three sets of ring camera arrays, full-surface images of the piston's upper ball head area, middle rod body area and lower conical surface area are acquired to obtain the corresponding full-surface images of the piston; S2: Intelligent Defect Recognition: The full surface image of the piston is used as the input of a multi-scale feature fusion deep learning model, and the corresponding defect is obtained as the output. S3: Defect Mapping Determination: Based on the corresponding camera coefficients in the ring camera array, the identified defects are mapped to the three-dimensional digital model of the piston to obtain the corresponding defect detection report.
2. The intelligent detection method for automotive air conditioning compressor piston surface defects based on visual analysis according to claim 1, characterized in that, The corresponding full surface image of the piston is obtained, including: S1.1: Camera distribution: Macro cameras are set at the entrance and exit of the closed imaging box through camera mounting rings, and multiple industrial cameras are set in the middle of the closed imaging box, and the pitch angle of the macro cameras is greater than that of the industrial cameras. S1.2: Image Acquisition: An illumination unit is set on the camera mounting ring, and through information interaction with the central controller, under the illumination of the illumination unit, the macro camera acquires the full surface image of the upper ball head area and the full surface image of the lower cone area of the piston, and the industrial camera acquires the full surface image of the middle rod area of the piston. S1.3: Image Enhancement: By using a structured light 3D scanning algorithm and multi-focus depth-of-field synthesis, image enhancement processing is performed on the full surface images of the upper ball head region and the lower cone region of the piston to obtain a full surface image with enhanced details.
3. The intelligent detection method for piston surface defects in automotive air conditioning compressors based on visual analysis according to claim 2, characterized in that, The macro camera has a pitch angle greater than 70°, while the industrial camera has a pitch angle set to 10°-30°.
4. The intelligent detection method for surface defects of automotive air conditioning compressor pistons based on visual analysis according to claim 2, characterized in that, The illumination unit corresponding to the macro camera is equipped with a ring light or a strip light, and the illumination unit corresponding to the industrial camera is equipped with a dome light or a coaxial light, and the ring light or strip light is set on the same side or opposite side of the macro camera.
5. The intelligent detection method for piston surface defects in automotive air conditioning compressors based on visual analysis according to claim 2, characterized in that, The obtained full-surface image with enhanced details includes: S1.3.1: Multi-focus fusion: The shooting position of the macro camera is adjusted by a piezoelectric ceramic actuator, and full-surface images corresponding to different positions are acquired according to the adjusted shooting position. At the same time, the full-surface images corresponding to different positions are aligned at the pixel level by an image registration algorithm to obtain the sharpness score of each pixel. The pixels corresponding to the maximum sharpness score are stitched together to obtain the corresponding fused full-surface image. S1.3.2: Structured light reconstruction: During the macro camera shooting process, a light spot pattern is projected, a full surface image with deformed light spots is acquired, and the gray value corresponding to each pixel in the full surface image with deformed light spots is determined. At the same time, based on the gray value of each pixel in different full surface images with deformed light spots, the corresponding phase value is determined, the corresponding three-dimensional spatial coordinates are obtained, and the corresponding three-dimensional point cloud map is constructed based on the three-dimensional spatial coordinates. S1.3.3: Image fusion: The fused full-surface image and the three-dimensional point cloud image are fused at the pixel level to obtain a full-surface image with enhanced details.
6. The intelligent detection method for piston surface defects in automotive air conditioning compressors based on visual analysis according to claim 1, characterized in that, The output retrieves the corresponding identified defects, including: S2.1: Model Setup: A multi-scale feature fusion deep learning model is constructed by setting up an improved deep learning network with sequentially connected dual input terminals, backbone network, neck network, and detection head. S2.2: Model Prediction: Using dual input terminals, the full surface image of the piston's middle rod region and the full surface image with enhanced details are used as inputs to a multi-scale feature fusion deep learning model, and the corresponding defect identification results are obtained as outputs.
7. The intelligent detection method for piston surface defects in automotive air conditioning compressors based on visual analysis according to claim 6, characterized in that, The backbone network is provided with a high-resolution feature branch, and the neck network is provided with a high-resolution thinning branch. Both the high-resolution feature branch and the high-resolution thinning branch are used to process the full-surface image after detail enhancement.
8. The intelligent detection method for automotive air conditioning compressor piston surface defects based on visual analysis according to claim 1, characterized in that, The corresponding defect detection report is obtained, including: S3.1: Coordinate Determination: All cameras in the circular camera array are calibrated to obtain the intrinsic and extrinsic parameters corresponding to each camera. Based on the intrinsic and extrinsic parameters corresponding to the cameras, the preliminary three-dimensional coordinates of each defect information under different cameras are determined. At the same time, the preliminary three-dimensional coordinates of each defect information are combined by triangulation to determine the final three-dimensional coordinates corresponding to each defect information. S3.2: Data Deduplication: Based on the final three-dimensional coordinates corresponding to each defect information, the Euclidean distance between two adjacent defect information is obtained. Based on the comparison between the Euclidean distance and the preset fusion threshold, the defect identification results are deduplicated. The deduplicated defect identification results are then mapped in the piston three-dimensional digital model to obtain the corresponding defect detection report with three-dimensional coordinates.
9. The intelligent detection method for piston surface defects in automotive air conditioning compressors based on visual analysis according to claim 8, characterized in that, Based on the comparison between the Euclidean distance and the preset fusion threshold, the defect identification results are deduplicated, specifically as follows: When the Euclidean distance is greater than the preset fusion threshold, the corresponding defect information is not fused, and the final three-dimensional coordinates corresponding to the defect information are the predicted three-dimensional coordinates; otherwise, the two corresponding defect information are fused, and one defect information is retained among the two adjacent defect information, and the final three-dimensional coordinates corresponding to the retained defect information are the predicted three-dimensional coordinates.
10. A visual analysis-based intelligent detection system for piston surface defects in automotive air conditioning compressors, characterized in that, The intelligent detection method for piston surface defects in automotive air conditioning compressors based on visual analysis, as described in any one of claims 1-9, was used.