3D printing powder laying defect detection method and system based on two-stage cascade model

Abstract

The application discloses a 3D printing powder laying defect detection method and system based on a two-stage cascade model, which comprises the following steps: a classification model based on a YOLO architecture is used to quickly classify powder laying images, and the images are divided into three categories, namely, clear qualified, clear unqualified and fuzzy unqualified; for the fuzzy unqualified images, an image preprocessing procedure containing Gaussian filtering, contrast enhancement and smoothing processing is performed to enlarge the small gray difference between the waves and the background; then the enhanced images are sent to a target detection model for wave area positioning detection; finally, a higher confidence threshold is set to filter the detection results, and the over-detection caused by the preprocessing noise amplification is effectively inhibited. The application significantly reduces the over-detection rate under the premise of ensuring the defect detection rate, and provides defect area positioning information, and the detection accuracy and real-time performance are taken into account.

Description

Technical Field

[0001] This invention relates to the field of additive manufacturing technology, and more specifically, to a method and system for detecting powder spread defects in 3D printing based on a two-stage cascade model. Background Technology

[0002] 3D printing (additive manufacturing) technology constructs three-dimensional solid parts by depositing materials layer by layer, and has been widely used in aerospace, medical devices, and automotive manufacturing. In powder bed fusion (PBF) 3D printing processes, powder bed quality is one of the key factors affecting the performance of the printed parts. The uniformity of powder bed placement in each layer directly determines key indicators such as interlayer bonding strength, density, and surface roughness.

[0003] During the powder spreading process, various defects may appear on the powder-spreading surface due to factors such as wear of the spreading scraper, changes in powder flowability, and improper spreading speed. Among these, wavy lines (also known as textures or ripple defects) are the typical manifestation and gold standard characteristic of powder spreading defects. Wavy lines appear as periodic stripe-like textures on the powder-spreading surface. They are caused by uneven force on the powder during the movement of the spreading mechanism, resulting in periodic fluctuations in the powder layer thickness.

[0004] In existing technologies, the following methods are mainly used for detecting powder spreading defects: 1. Manual visual inspection. Operators manually determine the presence of defects by observing the powder-coated surface through monitoring cameras or on-site observation. This method relies on personnel experience, has inconsistent inspection standards, cannot meet the needs of real-time online inspection, and is inefficient.

[0005] 2. Automated detection based on a single classification model. Deep learning classification models (such as VGG, ResNet, etc.) are used to perform binary classification (qualified / unqualified) on the powder-spreading images. While this method can achieve automated detection, it has the following technical limitations:

[0006] (1) High over-detection rate: Due to the weak wavy texture feature in the toner-laying image (the gray level difference with the background is usually only 3-10 gray levels), the classification model is prone to oversensitivity when processing boundary samples, misjudging normal images as NG (unacceptable), resulting in a high over-detection rate. Over-detection will cause unnecessary printing interruptions or layer re-laying, which will seriously affect printing efficiency.

[0007] (2) Lack of positioning capability: The classification model only outputs the judgment results at the whole map level, and cannot provide the location information of the defect area, which is not conducive to subsequent defect analysis and process optimization.

[0008] (3) Insufficient utilization of features: The single classification model does not specifically process the specific texture features of the wave pattern, making it difficult to distinguish the subtle differences between real wave patterns and normal surface texture fluctuations.

[0009] 3. Directly using the target detection model. The target detection model is directly run on all powder-spreading images to locate the wavy pattern region. Although this method can provide defect location information, the detection rate is low due to the extremely weak wavy pattern features, and the computational cost of detecting all images is high, making it difficult to meet real-time requirements.

[0010] In summary, there is an urgent need for a powder-laying defect detection solution that can effectively reduce the over-detection rate while ensuring the detection rate, and at the same time take into account the detection efficiency. Summary of the Invention

[0011] To address at least one of the aforementioned technical problems, this invention proposes a method and system for detecting powder spreading defects in 3D printing based on a two-stage cascade model.

[0012] The first aspect of this invention provides a method for detecting powder spread defects in 3D printing based on a two-stage cascade model, comprising the following steps: S1, acquire an image of the powder-spread surface for 3D printing; S2, the image is input into the first-stage classification model for classification. The classification model is based on the YOLO architecture and outputs a judgment result of clearly qualified, clearly unqualified, or fuzzy unqualified. Among them, images that are clearly qualified and clearly unqualified directly output the judgment result, while images that are fuzzy unqualified enter the second stage. S3 performs image preprocessing on blurry and unqualified images, including Gaussian filtering, contrast enhancement, and smoothing, to amplify the grayscale difference between the wavy lines and the background. S4. Input the preprocessed image into the second-stage target detection model to detect the wavy texture region and output the detection box and confidence score. S5. Filter the detection results based on confidence level. If the detection confidence level is greater than or equal to the preset confidence level threshold, the result is considered unqualified; otherwise, it is considered qualified.

[0013] In a preferred embodiment of the present invention, in step S2, the determination logic of the first-stage classification model is as follows: In step S2, the determination logic of the first-stage classification model is as follows: When the probability of passing is greater than or equal to the first threshold T1 (usually set to 0.5), it is judged as clearly passing; when the probability of failing is greater than or equal to the second threshold T2 (usually set to 0.85), it is judged as clearly failing. T1 and T2 are set by the user. When the probability of non-compliance is within the judgment interval, it is judged as fuzzy non-compliance; In a preferred embodiment of the present invention, the image preprocessing in step S3 includes the following sub-steps: S31, Gaussian filtering is performed on the blurred and unqualified image using a Gaussian kernel to remove high-frequency noise and retain the low-frequency texture features of the wavy pattern. The standard deviation σ of the Gaussian kernel ranges from 1 to 3. S32 performs global contrast enhancement processing on the Gaussian filtered image, using histogram equalization to improve image contrast and amplify grayscale jumps at the edges of wavy lines. S33 performs smoothing processing on the contrast-enhanced image, using mean filtering to homogenize the gray levels in the background area and eliminate local gray-level fluctuations that may be introduced by contrast enhancement.

[0014] In a preferred embodiment of the present invention, the Gaussian pyramid downsampling step is further included before the Gaussian filtering: when the maximum side length of the image is greater than a set threshold, the image is downsampled after performing Gaussian kernel convolution filtering, and the downsampling operation is repeated until the maximum side length of the image meets the size requirements.

[0015] In a preferred embodiment of the present invention, the preset confidence threshold Th in step S5 is 0.75. The threshold is higher than the confidence threshold of conventional target detection and is used to suppress overdetection caused by noise amplification due to preprocessing enhancement.

[0016] In a preferred embodiment of the present invention, the wavy pattern is a periodic stripe texture that appears on the powder-coated surface and is the gold standard feature of powder-coated defects. The gray level difference between the wavy pattern and the background is 3-10 gray levels before preprocessing and is amplified to 30-60 gray levels after preprocessing.

[0017] In a preferred embodiment of the present invention, both the first-stage classification model and the second-stage object detection model are based on the YOLO architecture. The classification model uses a classification head to output class probabilities, and the object detection model uses a detection head to output detection box coordinates and confidence scores. The training data of the second-stage object detection model undergoes the same preprocessing procedure as the inference stage to ensure the consistency of feature distribution between training and inference.

[0018] A second aspect of this invention provides a 3D printing powder spread defect detection system based on a two-stage cascade model, applied to a 3D printing powder spread defect detection method based on a two-stage cascade model, comprising: The image acquisition module is used to acquire images of the powder-spread surface in 3D printing. The first-stage classification module deploys a classification model based on the YOLO architecture to quickly classify the input image and output a judgment result of clearly qualified, clearly unqualified, or fuzzy unqualified. The image preprocessing module performs Gaussian filtering, contrast enhancement, and smoothing on blurry and unqualified images, and amplifies the wavy texture features; The second-stage detection module deploys the YOLO target detection model to locate and detect wavy areas in the enhanced image, and outputs detection boxes and confidence scores. The confidence filtering module filters and judges the detection results based on a preset confidence threshold. The results output module feeds back the detection results to the 3D printer control system.

[0019] In the image preprocessing module, the standard deviation σ of the Gaussian kernel for Gaussian filtering is 1, and the kernel size is 5×5. Histogram equalization and 5×5 mean filtering are performed sequentially.

[0020] The technical solution of the present invention has the following advantages compared with the prior art: The first-stage classification model quickly filters out most of the clear samples, sending only blurry, out-of-memory (NG) images to the second stage, balancing detection accuracy and real-time performance.

[0021] By employing a combined preprocessing strategy of pyramid downsampling, contrast enhancement, and smoothing, previously invisible weak features are amplified to a detectable range.

[0022] In the second stage of detection, a higher confidence threshold is set to effectively eliminate false positives caused by preprocessing amplification noise, thereby significantly reducing the over-detection rate.

[0023] The second-stage target detection model outputs the location and extent of the defect area, providing data support for subsequent process parameter adjustments and defect tracing. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, some of the drawings in the following description are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is an overall flowchart of the two-stage detection algorithm for 3D printing powder spread defects provided in an embodiment of the present invention; Figure 2 This is a diagram of the two-stage cascaded detection model architecture provided in an embodiment of the present invention; Figure 3 The image preprocessing flow and wavy texture enhancement effect comparison diagram provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the wavy texture feature of the powder-coated surface provided in an embodiment of the present invention. The left side is the original image, and the right side is the image after enhancement processing. Detailed Implementation

[0026] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0028] Example 1 See Figures 1-4 As shown, this invention proposes a method for detecting powder spread defects in 3D printing based on a two-stage cascade model, comprising the following steps: Step S1: Image acquisition.

[0029] An industrial area scan camera with a resolution of at least 5 megapixels is installed directly above the powder bed of the 3D printer, illuminated by a ring light source or coaxial light source. After each layer of powder is laid, the camera is triggered to capture an image of the powder-spread surface. The image resolution is 1500×1500 pixels, in grayscale format. The captured image is transmitted to an inspection server via network as input for subsequent inspection processes.

[0030] Step S2: First-stage classification and detection.

[0031] The collected powder-spreading images are input into the first-stage classification model. In this embodiment, the classification model adopts the YOLOv8-cls architecture, with CSPDarknet as the backbone network for feature extraction, and the classification head outputs the probability values ​​of two categories: OK (qualified) and NG (unqualified).

[0032] The training dataset for the classification model contains 10,000 images of powder spreading, including 6,000 OK samples and 4,000 NG samples, all manually labeled. Training employs a transfer learning strategy, fine-tuning the pre-trained weights from COCO with a learning rate of 0.001, a batch size of 32, and input images of 1024×1024 pixels. Training lasts for 200 epochs.

[0033] The output decision logic of the classification model is as follows: If the OK probability is ≥0.5, it is determined to be OK and a qualified result is output directly; If the probability of NG is ≥0.85, it is determined as a clear NG, and the unqualified result is output directly; If the NG probability ∈ [0.5, 0.85), it is judged as fuzzy NG and enters the second stage of detection.

[0034] Using the above strategy, most clear samples can be judged in the first stage, and only about 15%-25% of blurry NG images need to enter the second stage, effectively ensuring the overall detection efficiency.

[0035] like Figure 3 As shown, for the blurred NG images entering the second stage, a targeted image preprocessing workflow is performed: Step S31: Pyramid downsampling and Gaussian filtering.

[0036] Gaussian pyramid downsampling (pyrDown) is used to perform Gaussian filtering and downsampling on the image to be processed. This operation first uses a two-dimensional Gaussian kernel to perform convolution filtering on the image to be processed. The Gaussian kernel function is defined as: G(x,y)=(1 / (2πσ²))×e^(-(x²+y²) / (2σ²)); Where σ is the standard deviation of the Gaussian kernel, and a 5×5 Gaussian kernel (σ≈1) is used in this embodiment. The purpose of Gaussian filtering is to remove high-frequency noise (such as sensor noise, ambient light fluctuations, etc.) from the image while preserving the low-frequency periodic texture features of the wavy lines. After filtering, the image is downsampled by a factor of 2 to reduce the image size and decrease the computational load of subsequent processing. When the maximum side length of the image is still greater than a set threshold, the downsampling operation is repeated until the size requirement is met. By using pyramid downsampling, high-frequency details are gradually discarded, while more stable low-frequency information is preserved. Through multiple Gaussian filters, noise is effectively suppressed, preserving more detail information than direct downsampling.

[0037] Step S32: Contrast enhancement.

[0038] Histogram equalization is applied to the Gaussian filtered and downsampled image to enhance contrast. Histogram equalization expands the dynamic range of grayscale values ​​by changing the grayscale histogram of the original image from a concentrated distribution to a uniform distribution.

[0039] Step S33: Smoothing process.

[0040] The contrast-enhanced image was smoothed using a box filter with a kernel size of 5×5. The box filter replaces the center pixel value with the mean value of all pixels within a local window, further homogenizing the grayscale distribution in the background area and eliminating noise introduced by histogram equalization. After smoothing, the background grayscale in the normally powdered areas becomes more uniform, while the wavy areas, due to the significantly enhanced grayscale contrast after contrast enhancement, retain their periodic grayscale fluctuation characteristics.

[0041] Explanation of the principle of preprocessing enhancement: The essence of the wavy pattern is the periodic change in grayscale caused by the periodic fluctuation of the powder layer thickness on the powder-coated surface under illumination. In the original image, this grayscale change is extremely small (the background grayscale is about 128, the peak value of the wavy pattern is about 131, and the grayscale difference is only about 3), which is difficult for the human eye and conventional detection models to distinguish.

[0042] First, pyramid downsampling is used to reduce the image size, decreasing the computational burden while better preserving the original image information. Next, Gaussian filtering is used to remove high-frequency noise. Then, histogram equalization remaps the global grayscale range, stretching and amplifying minute grayscale jumps at the edges of the wavy lines (magnifying the grayscale difference from approximately 3 levels to 30-60 levels), while the grayscale variations in the background area are compressed due to its concentrated grayscale and narrow histogram. Finally, box filtering performs mean smoothing on the enhanced image, eliminating the amplification effect of local noise that may be introduced during enhancement, further homogenizing the background grayscale distribution. Ultimately, this achieves the enhancement effect of "uniform background and enhanced wavy lines."

[0043] like Figure 2 As shown, the preprocessed and enhanced image is input into the second-stage object detection model. In this embodiment, the object detection model adopts the YOLOv8-det architecture, with CSPDarknet as the backbone network and FPN+PAN as the neck structure. The detection head outputs the detection box coordinates and confidence scores.

[0044] The training dataset for the object detection model contains 3,000 images labeled with wavy region detection boxes. All training images undergo the same preprocessing steps as the inference stage (Gaussian filtering + histogram equalization + smoothing) to ensure consistency in feature distribution between training and inference. Training uses COCO pre-trained weights with a learning rate of 0.001, a batch size of 16, a model input size of 1024×1024, and a training duration of 300 epochs.

[0045] The output of the detection model is a number of detection boxes. Each detection box contains: bounding box coordinates (x, y, w, h), category (wavy pattern), and confidence score.

[0046] The preset confidence threshold filtering specifically includes: Because preprocessing enhancement, while amplifying wavy texture features, may also amplify fluctuations in certain normal textures, it can cause the detection model to falsely detect non-defect areas. To suppress such overdetection, this invention sets a high threshold in the confidence filtering stage of the second-stage detection results.

[0047] Specifically, the confidence threshold for conventional target detection is typically set to 0.25-0.5, while this invention sets the threshold Th to 0.7-0.85 (Th=0.75 in this embodiment). Only when the highest confidence level of the detection output satisfies: When max(confidence_i)≥Th, i=1,2,...,N, the image is deemed unacceptable (NG); otherwise, it is deemed acceptable (OK).

[0048] The rationale behind this high-threshold strategy is that, after preprocessing enhancement, the true wavy texture features are significantly amplified, and the confidence level of the detection model for true defects can typically reach above 0.8; while the confidence level for false detections caused by noise amplification is generally between 0.4 and 0.65. By setting Th=0.75, true defects and noise false detections can be effectively distinguished, significantly reducing the over-detection rate while ensuring the detection rate.

[0049] Threshold selection criteria: In the verification experiment of this embodiment, 500 test images (120 real NG images and 380 OK images) were detected, and the performance at different thresholds is as follows:

[0050] As shown in the table above, when Th=0.75, the F1-Score reaches the optimal level of 0.931, and the over-detection rate decreases from 18.4% of the conventional threshold to 3.4%, a reduction of 81.5%, while the detection rate remains at a high level of 93.3%.

[0051] This invention also provides a 3D printing powder spread defect detection system based on a two-stage cascade model, comprising: Image acquisition module: including industrial camera, light source and image transmission unit, used to acquire images of the powder surface after each layer of powder is spread and transmit the images to the detection server; The first-stage classification module deploys a classification model based on the YOLO architecture (YOLOv8-cls in this embodiment) to quickly classify the input image and output a clear OK, clear NG, or fuzzy NG judgment. Images with clear OK and clear NG are directly output as results, while fuzzy NG images are input to subsequent modules. Image preprocessing module: Performs Gaussian filtering (with optional downsampling), histogram equalization, contrast enhancement, and mean smoothing on blurred NG images to amplify wavy texture features; The second-stage detection module deploys the target detection model (YOLOv8-det in this embodiment) to locate and detect the wavy texture region in the enhanced image, and outputs the detection box and confidence score. Confidence filtering module: Filters the test results according to the preset confidence threshold (Th=0.7-0.85) and outputs the final pass or fail judgment; Results output module: Feeds the test results back to the 3D printer control system and supports subsequent operations such as defect alarms, printing interruptions, or layer re-layouts.

[0052] The above modules work together to form a complete automatic detection process for powder spreading defects. The first-stage classification module quickly filters out most of the clear samples, inputting only the blurry NG images into subsequent processing, balancing detection accuracy and real-time performance; the image preprocessing module specifically enhances the weak features of wavy patterns, breaking through the bottleneck of feature detection; the confidence filtering module effectively suppresses over-detection through a high-threshold strategy, achieving the optimal balance between detection rate and over-detection rate.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to the above embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0055] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting powder spread defects in 3D printing based on a two-stage cascade model, characterized in that, Includes the following steps: S1, acquire an image of the powder-spread surface for 3D printing; S2, the image is input into the first-stage classification model for classification. The classification model is based on the YOLO architecture and outputs a judgment result of clearly qualified, clearly unqualified, or fuzzy unqualified. Among them, images that are clearly qualified and clearly unqualified directly output the judgment result, while images that are fuzzy unqualified enter the second stage. S3 performs image preprocessing on blurry and unqualified images, including Gaussian filtering, contrast enhancement, and smoothing, to amplify the grayscale difference between the wavy lines and the background. S4. Input the preprocessed image into the second-stage target detection model to detect the wavy texture region and output the detection box and confidence score. S5. Filter the detection results based on confidence level. If the detection confidence level is greater than or equal to the preset confidence level threshold, the result is considered unqualified; otherwise, it is considered qualified.

2. The 3D printing powder spreading defect detection method based on a two-stage cascade model according to claim 1, characterized in that, In step S2, the decision logic of the first-stage classification model is as follows: When the probability of passing is greater than or equal to the first threshold T1, it is determined to be clearly passing; when the probability of failing is greater than or equal to the second threshold T2, it is determined to be clearly failing. When the probability of non-compliance is within the judgment interval, it is judged as fuzzy non-compliance.

3. The 3D printing powder spreading defect detection method based on a two-stage cascade model according to claim 1, characterized in that, The image preprocessing in step S3 includes the following sub-steps: S31, Gaussian filtering is performed on the blurred and unqualified image using a Gaussian kernel to remove high-frequency noise and retain the low-frequency texture features of the wavy pattern. The standard deviation σ of the Gaussian kernel ranges from 1 to 3. S32 performs global contrast enhancement processing on the Gaussian filtered image, using histogram equalization to improve image contrast and amplify grayscale jumps at the edges of wavy lines. S33 performs smoothing processing on the contrast-enhanced image, using mean filtering to homogenize the gray levels in the background area and eliminate local gray-level fluctuations that may be introduced by contrast enhancement.

4. The 3D printing powder spreading defect detection method based on a two-stage cascade model according to claim 1, characterized in that, The Gaussian filtering process includes a Gaussian pyramid downsampling step: when the maximum side length of the image is greater than a set threshold, the image is downsampled after Gaussian kernel convolution filtering, and the downsampling operation is repeated until the maximum side length of the image meets the size requirements.

5. The 3D printing powder spreading defect detection method based on a two-stage cascade model according to claim 1, characterized in that, The preset confidence threshold Th in step S5 has a value range of 0.

75. This threshold is higher than the confidence threshold for conventional target detection and is used to suppress over-detection caused by noise amplification due to preprocessing enhancement.

6. The 3D printing powder spreading defect detection method based on a two-stage cascade model according to claim 1, characterized in that, The wavy pattern is a periodic stripe texture that appears on the powder-coated surface. It is the gold standard feature of powder-coated defects. The gray level difference between the wavy pattern and the background is 3-10 gray levels before preprocessing, and is amplified to 30-60 gray levels after preprocessing.

7. The 3D printing powder spreading defect detection method based on a two-stage cascade model according to claim 1, characterized in that, Both the first-stage classification model and the second-stage object detection model are based on the YOLO architecture. The classification model uses a classification head to output class probabilities, and the object detection model uses a detection head to output detection box coordinates and confidence scores. The training data of the second-stage object detection model undergoes the same preprocessing process as the inference stage to ensure the consistency of feature distribution between training and inference.

8. A 3D printing powder spread defect detection system based on a two-stage cascade model, applied to the 3D printing powder spread defect detection method based on a two-stage cascade model as described in any one of claims 1-7, characterized in that, include: The image acquisition module is used to acquire images of the powder-spread surface in 3D printing. The first-stage classification module deploys a classification model based on the YOLO architecture to quickly classify the input image and output a judgment result of clearly qualified, clearly unqualified, or fuzzy unqualified. The image preprocessing module performs Gaussian filtering, contrast enhancement, and smoothing on blurry and unqualified images, and amplifies the wavy texture features; The second-stage detection module deploys a target detection model to locate and detect wavy areas in the enhanced image, and outputs detection boxes and confidence scores. The confidence filtering module filters and judges the detection results based on the confidence threshold. The results output module feeds back the detection results to the 3D printer control system.

9. A 3D printing powder spreading defect detection system based on a two-stage cascade model according to claim 8, characterized in that, In the image preprocessing module, the standard deviation σ of the Gaussian kernel for Gaussian filtering is 1, and the kernel size is 5×5. Histogram equalization and 5×5 mean filtering are performed sequentially.

Images