Glass bottle bottom defect detection method and device

Abstract

The invention relates to the technical field of glass bottle detection, and discloses a glass bottle bottom defect detection method which comprises the following steps: building a detection platform, and adopting a multi-view imaging module of an annular LED and oblique light supplement combined light source, a vertical camera and 3-4 oblique cameras; after the to-be-detected glass bottle is positioned, synchronously collecting a bottle bottom full-view image set; graying, adaptive median filtering, CLAHE histogram equalization and Otsu binarization preprocessing are carried out on the image; extracting geometric and textural features of the image, inputting the geometric and textural features into an improved YOLOv5 model for identification and classification, and judging defects by combining with multi-view result fusion; and sorting the glass bottles according to a detection result and generating a traceable report. The identification rate of tiny defects is larger than or equal to 99%, the detection time of a single bottle is smaller than or equal to 0.5 second, and the method is compatible with glass bottles with the diameter of 30-100 mm, is suitable for large-scale quality control in the fields of food and beverage, medicine packaging and the like, and has extremely high practical value.

Description

Technical Field

[0001] This invention relates to the field of glass bottle inspection, and in particular to a method and apparatus for detecting defects at the bottom of glass bottles. Background Technology

[0002] As a core packaging container in the food, beverage, and pharmaceutical industries, the quality of glass bottles directly affects product safety and consumer health. During the glass bottle manufacturing process, the bottom of the bottle is prone to defects such as cracks, bubbles, sand particles, dents, and bulges due to the complex molding process (e.g., uneven mold temperature, poor glass flow). Among these, tiny cracks (width < 0.1 mm) may expand and cause leakage during liquid filling or sterilization, while fine sand particles (diameter < 0.1 mm) can detach and mix into the contents, posing a risk of food safety or pharmaceutical contamination. Therefore, bottom defect detection is a crucial aspect of glass bottle quality control. Existing glass bottle bottom defect detection technologies have significant shortcomings: Manual inspection is difficult to meet the needs of large-scale production: relying on operators' visual observation, the inspection efficiency is only 20-30 bottles per minute, which cannot keep up with the pace of more than 100 bottles per minute in modern production lines; moreover, manual judgment is affected by fatigue and experience differences, and the rate of missed detection of minor defects is as high as 10%-15%, while subjective misjudgment exists (such as misjudging the texture on the bottom of the bottle as a crack), making it difficult to guarantee quality stability. Traditional machine vision inspection methods suffer from insufficient accuracy and anti-interference capabilities: they often use a single light source (such as vertical white light) and simple image processing algorithms (such as fixed threshold segmentation). Reflections on the bottle bottom and mold textures can easily create interference noise, making it difficult to identify minute defects (such as 0.02mm cracks). Furthermore, they often use single-view imaging, which cannot cover the area where the bottle bottom edge meets the side wall (this area is prone to hidden cracks). The false negative rate is generally 5%-8%, which is difficult to meet the high-precision requirements of pharmaceutical packaging and other scenarios. Existing technologies suffer from poor versatility and scalability: Traditional testing equipment is mostly designed for specific glass bottle sizes. When changing bottle types, the light source angle and camera position need to be readjusted, with a debugging cycle of up to 1-2 days. Furthermore, it is not compatible with small-sized medicine bottles with a diameter of less than 30mm and large-capacity food bottles with a diameter of more than 100mm. At the same time, there is a lack of defect data traceability and process optimization linkage mechanisms, which cannot provide effective feedback to upstream production (such as mold maintenance and glass melt temperature control), resulting in repeated defects. In summary, existing technologies cannot simultaneously achieve high detection accuracy, efficiency, and versatility; therefore, this invention proposes a method and apparatus for detecting defects at the bottom of glass bottles to solve the aforementioned problems. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides a method and apparatus for detecting defects at the bottom of glass bottles.

[0004] The present invention provides a method and apparatus for detecting defects at the bottom of glass bottles, which adopts the following technical solution: A method for detecting defects at the bottom of a glass bottle includes the following steps: (1) Inspection preparation: Build an inspection platform including a conveying mechanism, positioning device, multi-view optical imaging module, light source module, image acquisition card and industrial computer. The light source module adopts a combination of ring LED light source and oblique supplementary light source. The multi-view optical imaging module includes 1 vertical camera and 3-4 circumferentially distributed tilting cameras. (2) Image acquisition: The glass bottle to be tested is sent to the testing station by the conveying mechanism. After the positioning device is fixed, the light source module is turned on and the multi-view optical imaging module simultaneously acquires a full-view image set of the bottom of the bottle. (3) Image preprocessing: The image set is subjected to grayscale conversion, adaptive median filtering, histogram equalization and binarization; (4) Defect identification and classification: Extract the geometric and texture features of the preprocessed image, input them into the pre-trained deep learning model for identification and classification, and combine the results from multiple perspectives to determine whether the defect exists; (5) Results output: Based on the identification results, the glass bottles are sorted into qualified or unqualified channels and an inspection report is generated.

[0005] Preferably, in step (1), the illumination angle of the oblique supplementary light source is 30°-60°, and the image resolution of the multi-view optical imaging module is not less than 2048×1536 pixels.

[0006] Preferably, in step (2), the positioning device fixes the glass bottle by mechanical limiting and vacuum adsorption, and the distance between the bottom of the bottle and the imaging module is stabilized at 50-80mm.

[0007] Preferably, in step (3), adaptive median filtering is used to remove noise, the CLAHE algorithm is used to enhance image contrast, and the Otsu algorithm is used for binarization.

[0008] Preferably, in step (4), the deep learning model is an improved YOLOv5 model with an added CBAM attention mechanism module. When making multi-view fusion judgments, the same defect must be identified and classified consistently in at least two images from different viewpoints.

[0009] A glass bottle bottom defect detection device includes a conveying mechanism, a light source module is arranged directly above the detection position of the conveying mechanism, a multi-view optical imaging module is arranged directly below it, and positioning mechanisms are arranged on both sides of the detection position of the conveying mechanism for positioning the glass bottle.

[0010] In summary, the present invention has at least one of the following beneficial technical effects: High detection accuracy: It adopts a combination of ring LED and oblique supplementary light source to eliminate interference, and is equipped with an improved YOLOv5 model with CBAM attention module and multi-view fusion judgment. It has an identification rate of over 99% for cracks ≥0.02mm and sand particles ≥0.05mm, with a false negative rate of ≤0.5% and a false positive rate of ≤0.3%, meeting the requirements of high-demand scenarios.

[0011] High testing efficiency: Fully automated operation, single bottle testing ≤0.5 seconds, 120-150 bottles per minute, suitable for production lines with an annual output of tens of millions of pieces; bottle type switching can be completed within 5 minutes, compatible with 30-100mm glass bottles, reducing equipment costs.

[0012] High stability and traceability: Standardized calibration procedures and anomaly handling mechanisms ensure stability; test data is linked to bottle labeling, supporting traceability for more than 3 years, complying with GMP standards, and facilitating process optimization. Attached Figure Description

[0013] Figure 1 This is a structural schematic diagram of an embodiment of the invention.

[0014] Explanation of reference numerals in the attached drawings: 1. Conveying mechanism; 2. Positioning mechanism; 3. Multi-view optical imaging module; 4. Light source module. Detailed Implementation

[0015] The following is in conjunction with the appendix Figure 1 The present invention will be described in further detail below. Example

[0016] I. Basic Implementation Scenario: General Glass Bottle Testing in the Food and Beverage Industry 1.1 Configuration of core equipment for the testing platform (including model and parameters) Device Module Specific configuration Selection criteria Conveying mechanism Belt conveyor (model: Dorner 2200 Series), 150mm wide, adjustable conveying speed 0.5-0.8m / s. Food-grade PU belt, wear-resistant and easy to clean, speed suitable for a production pace of 100-150 bottles per minute. Positioning device Mechanical limit block (stainless steel, positioning accuracy ±0.1mm) + vacuum suction cup (30mm diameter, suction pressure -0.06~-0.08MPa) Dual positioning ensures a stable distance between the bottle bottom and the imaging module, while vacuum adsorption prevents image blurring caused by bottle movement. Light source module Ring-shaped LED light source (model: Keyence CA-DRW3), power 50W, wavelength 550nm (green light); 4 angled supplementary LEDs (model: Osram SFH 4770), each with a power of 15W, illumination angle 45°, evenly distributed circumferentially. Green light has weak penetrating power, which can reduce reflections from impurities on the bottle bottom; 45° angled supplemental lighting can eliminate shadows on the bottle bottom texture and highlight minor defects. Multi-view imaging module One vertical camera (Basler acA2500-14uc, resolution 2592×1944 pixels, frame rate 30fps) + three tilting cameras (same model), tilt angle 30° (angle with vertical direction), lens focal length 16mm. 25-megapixel resolution meets the requirement of 0.02mm defect recognition, and a 30° tilt angle covers the bottom edge of the bottle (suitable for diameters of 30-80mm). Industrial Computer CPU: Intel Core i7-12700K, GPU: NVIDIA RTX 3090 (24GB VRAM), Memory: 32GB DDR5, Hard Drive: 2TB SSD The RTX 3090's computing power supports real-time inference of YOLOv5 models (single frame processing time ≤20ms), and SSDs ensure fast data read and write speeds. Image acquisition card Advantech PCIe-1747U, 4 Camera Link interfaces, 200MB / s sampling rate. Simultaneous acquisition of data from 4 cameras with no delay, avoiding misalignment of images from multiple perspectives. 1.2 Step-by-step operation process (including key parameter control) Step 1: Platform calibration (perform before powering on daily) Distance calibration: Place a standard calibration block (60mm thick) at the testing station, adjust the camera height to ensure that the distance between the vertical camera and the surface of the calibration block is 60mm (error ≤ ±0.5mm), and save the parameters using the calibration software built into the industrial computer; Light source calibration: Turn on the light source, take an image using a grayscale card (24 levels), and adjust the brightness of the ring light source (initial value 80%) and the brightness of the oblique fill light (initial value 60%) until the grayscale value deviation of each level of the grayscale card is ≤5, to ensure the uniformity of illumination; Model calibration: Input 100 standard glass bottle images with known defects (25 images each of cracks and bubbles, and 10 images each of sand, dents, and bulges), run the recognition model, and if the recognition accuracy is ≥99%, the calibration is complete; if it is lower than 99%, reload the pre-trained model weights (model file storage path: D:\Model\YOLOv5_CBAM_V1.2.pt).

[0017] Step 2: Loading and positioning the glass bottle to be tested Before the glass bottle is conveyed to the inspection station by the conveyor belt, the position of the bottle needs to be detected by the front sensor (model: BannerQ45VR3LP). When the sensor triggers the signal, the conveyor belt decelerates to 0.3m / s to ensure that the bottle enters the positioning area smoothly. The mechanical limit blocks clamp the bottle from both sides (clamping force 50N to prevent bottle deformation), and at the same time the vacuum suction cup starts adsorption. After adsorption time of 0.2 seconds, the conveyor belt stops and waits for image acquisition.

[0018] Step 3: Multi-view image acquisition (synchronization control) The industrial computer sends a trigger signal (signal delay ≤1ms) via the IO card to simultaneously turn on the light source and camera. Image acquisition is completed within 10ms after the light source is turned on (to avoid brightness fluctuations caused by light source preheating). The four captured images (one vertical and three tilted) are automatically stored in the specified folder (path: D:\Image\20251119\, naming format: bottle number_viewpoint_timestamp.bmp), and a capture log (containing information such as bottle number, capture time, and image resolution) is generated at the same time.

[0019] Step 4: Image preprocessing (parameters can be dynamically adjusted) Grayscale conversion: A weighted average method (grayscale value = 0.299R + 0.587G + 0.114B) is used to avoid the loss of defect information in a single channel; Adaptive median filtering: The initial size of the filtering window is 3×3. When the difference between the pixel gray value and the surrounding area is detected to be ≥20, the window automatically expands to 5×7 to ensure that the defect edges (such as 0.02mm cracks) are preserved while removing noise. CLAHE histogram equalization: ClipLimit set to 2.0, grid size 8×8, to enhance low-contrast areas (such as the grayscale difference between fine sand grains and the bottom of a bottle). Otsu binarization: Automatically calculates the threshold (usually between 120-150). If the threshold exceeds this range, the system will automatically prompt "abnormal lighting" and the light source needs to be recalibrated.

[0020] Step 5: Defect Identification and Fusion Judgment (Logic Refinement) Feature extraction: Connectivity analysis is performed on the binarized image to extract features including: area (number of pixels), perimeter (number of pixels), circularity (4π×area / perimeter²), aspect ratio (length / width of the minimum bounding rectangle), energy value of the gray-level co-occurrence matrix (reflecting texture uniformity) and entropy value (reflecting texture complexity). Model inference: Input the feature vector into the improved YOLOv5 model (including the CBAM attention module), and the model outputs the defect type (crack / bubble / sand / dent / protrusion) and confidence score (0-1). A confidence score ≥ 0.8 is considered a valid identification. Multi-perspective fusion: If the same defect is identified in two or more images from different viewpoints, and the defect position deviation is ≤5 pixels, the type is consistent, and the confidence level is ≥0.9, then the defect is confirmed to exist, and the defect coordinates (XY coordinate system with the center of the bottle bottom as the origin) are recorded. If only one image identifies a suspected defect (confidence level 0.8-0.9), calculate the feature similarity of the defect to the corresponding region in other viewpoint images (using the cosine similarity algorithm). If the similarity is ≥0.85, the defect is confirmed; otherwise, it is marked as a "suspected false detection" and requires manual review (the review entry is in the "anomaly database" of the detection software).

[0021] Step 6: Result Output and Sorting (Linkage Control) Inspection report generation: The system automatically compiles defect information for each bottle and generates an Excel report (including bottle number, inspection time, defect type, quantity, location, and confidence level). The report is automatically backed up to the cloud (Alibaba Cloud OSS) every hour. Sorting control: If no defects are detected, the industrial computer sends a "qualified" signal to the pneumatic push rod (model: FestoDSNU-20-50-PA), and the push rod moves the glass bottle to the qualified channel (response time ≤ 0.3 seconds); if defects are detected, the push rod pushes the bottle to the unqualified channel, and at the same time, the buzzer (frequency 2000Hz) sounds an alarm, and the display shows the defect image and type, prompting the staff to clean up the unqualified products in time.

[0022] 1.3 Performance Verification Results In this scenario, 10,000 food-grade glass bottles (60mm diameter at the bottom, including 1,000 samples with artificial implant defects) were tested, and the results are as follows: Defect identification rate: Cracks 99.3%, Bubbles 99.5%, Sand particles 99.2%, Dents 99.1%, Protrusions 99.4%, Average identification rate 99.3%; False negative rate: 0.3% (only 3 tiny cracks of 0.02mm were not detected because they were located in the blind spot at the bottom edge of the bottle. This was subsequently optimized by adding an inclined camera). False positive rate: 0.2% (all 20 false positive samples were due to abnormal bottle bottom texture; after adjusting CLAHE's ClipLimit to 1.8, the false positive rate dropped to 0.1%). Single bottle inspection time: 0.4 seconds (including 0.1 seconds for image acquisition, 0.1 seconds for preprocessing, 0.1 seconds for model inference, and 0.1 seconds for fusion judgment and sorting), meeting the production requirement of 150 bottles per minute.

[0023] II. High-precision implementation scenario: Inspection of glass bottles for pharmaceutical packaging 2.1 Core Differentiation Configurations (Compared to Basic Scenario) Imaging module: Add one tilting camera (out of four), tilted at 45°, to cover the area where the bottom edge of the bottle meets the side wall (the bottom of medicine bottles is thinner and prone to cracking at the joint); the angle of the oblique supplementary light source is adjusted to 60° to enhance the light intensity in the side wall joint area; Model optimization: The training dataset of the improved YOLOv5 model has been expanded with 50,000 images of defective medicine bottles (including "pinhole" defect type), and the channel attention weight of the CBAM attention module has been increased to 1.2, enhancing the ability to identify small defects (such as 0.01mm pinholes). Positioning accuracy: The vacuum adsorption pressure is increased to -0.09MPa, and the positioning accuracy of the mechanical limit block is increased to ±0.05mm, ensuring that the distance error between the bottom of the bottle and the camera is ≤±0.2mm (medical bottles have higher requirements for detection accuracy).

[0024] 2.2 Key Operational Adjustments (Targeting the Needs of the Pharmaceutical Scenarios) Aseptic control: All parts of the testing platform that come into contact with the glass bottles (such as belts, limit blocks, and suction cups) are made of 316L stainless steel or food-grade silicone. They are wiped and disinfected with 75% alcohol after daily testing and sterilized with ultraviolet light (sterilization time 30 minutes) weekly. Defect threshold tightening: When making multi-view fusion judgments, the confidence threshold is increased to 0.95 and the feature similarity threshold is increased to 0.9 to avoid the rejection of qualified medicine bottles due to false detection (zero tolerance for defects in medicine packaging). Data traceability: The test data of each bottle (including original image, preprocessing results, and identification report) is associated with the QR code on the bottle, supporting traceability by scanning the code (traceability period ≥ 3 years, in compliance with GMP standards).

[0025] 2.3 Performance Verification Results The results of testing on 5000 pharmaceutical-grade glass vials (bottom diameter 30mm, including 500 implantation defect samples and 200 pinhole defects of 0.01mm) are as follows: Average defect identification rate: 99.5% (pinhole identification rate: 99.0%). False negative rate: 0.2% (only one pinhole was found at the junction of the bottle bottom and side wall, which was resolved by adjusting the camera angle to 50°). False detection rate: 0.15% (no qualified bottles were rejected due to false detection); Single-bottle testing time: 0.5 seconds (the time is slightly longer due to the addition of one camera and more stringent fusion judgment, but it still meets the pharmaceutical production pace of 120 bottles per minute).

[0026] III. Exception Handling Mechanism 3.1 Equipment Malfunction Handling Exception types Triggering conditions Processing flow Abnormal light source brightness If the grayscale card detects brightness fluctuations ≥10%, or if the Otsu threshold of 5 consecutive bottle images exceeds the range of 120-150, the result is incorrect. The system automatically shuts down the light source and restarts the light source module (restart time 3 seconds). If the problem persists, it will prompt "Replace light source" and switch to the backup light source (one backup light source needs to be configured in advance). Camera capture failed No images were captured for three consecutive bottles, or the image resolution was lower than 2048×1536 pixels. Check the Camera Link interface connection (the system will automatically pop up the interface detection interface). If the interface is normal, restart the camera (restart time is 5 seconds). If it still fails, mark it as "camera failure" and switch to the backup camera (one backup vertical camera needs to be configured in the basic scene). Vacuum adsorption failure The vacuum pressure sensor detects a pressure ≥ -0.05 MPa, or the bottle body shakes ≥ 1 mm after positioning. Turn off adsorption and check if the suction cup is clogged (the system will prompt "Clean suction cup"). After cleaning, try adsorption again. If it fails three times in a row, a "Vacuum pump failure" message will appear. Activate the mechanical clamping backup solution (add one set of mechanical grippers with a clamping force of 80N). 3.2 Handling of Abnormal Test Results Suspected Defect Review: When the system marks a suspected false positive (single-view recognition with a similarity of 0.8-0.85), the image is automatically stored in the "Review Library". Staff must complete the review within 10 minutes (the review interface supports image magnification up to 400% to compare defect features). The review results are updated to the inspection report simultaneously. Batch Defect Warning: When 10 consecutive bottles are found to have the same type of defect (such as bubbles), or the defect rate of a single batch (100 bottles) is ≥5%, the system will automatically trigger a "batch warning", suspend the conveyor belt, and prompt staff to check the upstream production process (such as glass melting temperature and mold status). At the same time, a "defect trend analysis chart" (statistically calculates the percentage of defect types by time dimension) to assist in production optimization.

[0027] IV. Implementation Precautions (To ensure long-term stable operation) 1. Routine maintenance: After each daily inspection, clean the camera lens (wipe with a lint-free cloth dampened with anhydrous ethanol), clean the conveyor belt (blow away impurities with a high-pressure air gun), and check the light source cover for stains; calibrate the camera distance and light source brightness weekly, and replace the vacuum suction cup monthly (to avoid aging that could lead to suction failure). 2. Environmental control: The testing workshop must maintain a constant temperature (20-25℃) and constant humidity (40%-60% relative humidity) to avoid temperature fluctuations that could reduce equipment accuracy; at the same time, it should be kept away from vibration sources (such as air compressors) to prevent the camera and light source from shifting positions; Personnel training: Operators need to master the basic calibration process (such as light source and distance calibration) and anomaly handling methods (such as troubleshooting camera acquisition failures). Technicians need to be familiar with the model update process (when a new defect type is added, 5,000 images need to be annotated, the model needs to be retrained and the weight file replaced).

[0028] It should be noted that, in this document, terms such as “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0029] 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 variations 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 claims and their equivalents.

Claims

1. A method for detecting defects at the bottom of a glass bottle, characterized in that: Includes the following steps: (1) Inspection preparation: Build an inspection platform including a conveying mechanism, positioning device, multi-view optical imaging module, light source module, image acquisition card and industrial computer. The light source module adopts a combination of ring LED light source and oblique supplementary light source. The multi-view optical imaging module includes 1 vertical camera and 3-4 circumferentially distributed tilting cameras. (2) Image acquisition: The glass bottle to be tested is sent to the testing station by the conveying mechanism. After the positioning device is fixed, the light source module is turned on and the multi-view optical imaging module simultaneously acquires a full-view image set of the bottom of the bottle. (3) Image preprocessing: The image set is subjected to grayscale conversion, adaptive median filtering, histogram equalization and binarization; (4) Defect identification and classification: Extract the geometric and texture features of the preprocessed image, input them into the pre-trained deep learning model for identification and classification, and combine the results from multiple perspectives to determine whether the defect exists; (5) Results output: Based on the identification results, the glass bottles are sorted into qualified or unqualified channels and an inspection report is generated.

2. The method for detecting defects at the bottom of a glass bottle according to claim 1, characterized in that: In step (1), the illumination angle of the oblique supplementary light source is 30°-60°, and the image resolution of the multi-view optical imaging module is not less than 2048×1536 pixels.

3. The method for detecting defects at the bottom of a glass bottle according to claim 1, characterized in that: In step (2), the positioning device fixes the glass bottle by mechanical limiting and vacuum adsorption, and the distance between the bottom of the bottle and the imaging module is stabilized at 50-80mm.

4. The method for detecting defects at the bottom of a glass bottle according to claim 1, characterized in that: In step (3), adaptive median filtering is used to remove noise, the CLAHE algorithm is used to enhance image contrast, and the Otsu algorithm is used for binarization.

5. The method for detecting defects at the bottom of a glass bottle according to claim 1, characterized in that: In step (4), the deep learning model is an improved YOLOv5 model with an added CBAM attention mechanism module. When making multi-view fusion judgments, the same defect must be identified and classified consistently in at least two images from different viewpoints.

6. A glass bottle bottom defect detection device according to any one of claims 1-5, characterized in that: It includes a conveying mechanism (1), a light source module (4) is set directly above the detection position of the conveying mechanism (1), a multi-view optical imaging module (3) is set directly below it, and a positioning mechanism (2) is set on both sides of the detection position of the conveying mechanism (1) for positioning the glass bottle.

Images