Glass and non-woven fabrics are important industrial materials, and their surface quality directly affects product performance and applications. However, due to material characteristics and the complexity of the production process, surface defect detection has always been a difficult problem in the industry. DaoAI provides an effective solution to this problem with its advanced technology.
Industry background and user scenarios: Glass and non-woven fabrics are widely used in many fields, such as construction, medical, and textile industries. In glass production, high-quality glass is often used for building curtain walls and automotive glass. Minor surface defects may affect the overall safety and aesthetics. In non-woven fabric production, it is commonly used for medical protective supplies and filter materials. Surface defects can reduce its filtering performance and protective effect. Therefore, accurate detection of surface defects in glass and non-woven fabrics is crucial. For production enterprises, they need an efficient and accurate detection solution that can improve product quality and reduce the defective rate while ensuring production efficiency.
Deep - Dive into Pain Points: Why It's Difficult
From a quantitative perspective, the defect occurrence rate of glass and non-woven fabrics is relatively low, usually around 1% -5%. This results in a scarcity of defect samples available for training. Moreover, these defect samples have diverse morphologies. For example, the sizes of bubbles, shapes of stones, and lengths and depths of scratches on the glass surface vary. The shapes of holes and the types and sizes of foreign objects in non-woven fabrics also differ greatly. In actual detection, due to the semi-transparent nature of glass, the contrast between surface defects and the background is weak. Statistics show that the contrast between defect signals and background signals may be less than 10%, making it difficult to clearly identify defects. Non - woven fabrics have a naturally disordered fiber distribution, which is a typical weak-texture surface. The difficulty of separating defects from the background is extremely high, with a texture complexity index of over 80 (out of 100).
The root causes of these difficulties lie in the material properties and production processes. The semi-transparency of glass makes it susceptible to background and reflection interference during detection. The refraction and reflection of light can blur the boundaries of defects, increasing the difficulty of identification. In the production process of non-woven fabrics, the random arrangement of fibers forms complex textures, and defects are often hidden in these textures, making it difficult to distinguish them from normal textures. Traditional detection methods are often ineffective in such complex situations.
Technical Principles
DaoAI uses a technology that combines few-shot learning and APDT positive-sample anomaly detection. In terms of the algorithm mechanism, it first uses sufficient good-product images to establish a normal baseline, which is equivalent to setting a normal standard template for the system. Then, it uses a very small number of defect samples to fine-tune the criteria, enabling the system to identify abnormal situations different from the normal baseline. In terms of imaging and hardware mechanisms, the system is equipped with high-precision image acquisition devices that can capture the subtle features of the glass and non-woven fabric surfaces, providing accurate data for subsequent analysis.
Compared with traditional supervised deep-learning methods, traditional methods require a large number of annotated samples for effective training. However, in the real-world situation of scarce defect samples, it is impossible to gather enough samples, and projects often stall due to insufficient data. DaoAI's method can go live with only a small number of defect samples, greatly reducing the requirement for data volume. At the same time, based on the normal baseline of good products, it can also detect unseen defects and has stronger generalization ability.
Typical Application Scenarios
- Glass bubble detection: In the glass-forming process, bubble defects may occur. During detection, the system identifies bubbles by analyzing the changes in grayscale values in the image. The difficulty lies in the fact that the semi-transparency of glass makes the grayscale values of bubbles close to the background, and the system needs to accurately separate bubble signals under weak-contrast conditions.
- Glass stone detection: Stones are usually formed during the melting process of glass raw materials. During the detection process, the system makes judgments based on the texture differences between stones and the surrounding glass. Since the shapes and sizes of stones vary, and the contrast with the glass is not high, accurately identifying stones is a major challenge.
- Glass scratch detection: Scratches are generally produced during the cutting and handling of glass. The system detects scratches by detecting line features in the image. However, the reflection on the glass surface may interfere with the identification of scratches, and the image needs to be denoised and enhanced.
- Non - woven fabric hole detection: Holes may appear during the forming and processing of non-woven fabrics. The system identifies holes by analyzing the texture continuity of the image. Due to the disordered fiber characteristics of non-woven fabrics, it is difficult to distinguish holes from normal textures.
- Non - woven fabric foreign-object detection: Foreign objects may be mixed in the raw materials or enter the non-woven fabric during the production process. During detection, the system identifies foreign objects based on the color and texture differences between foreign objects and non-woven fabrics. However, when the color and texture of foreign objects are similar to those of non-woven fabrics, the detection difficulty increases.
Implementation Case
A medium-sized material production enterprise produces both glass and non-woven fabric products. Previously, the enterprise used manual visual inspection for surface defect detection. However, due to the influence of semi-transparency and weak textures in manual inspection, inspectors were prone to fatigue and missed detections during long-term inspections, and the consistency was poor, resulting in a relatively high defective rate. After introducing DaoAI's detection solution, the implementation process was relatively smooth. First, the DaoAI team used the enterprise's existing small number of defect samples and sufficient good-product samples for model training. Then, the trained model was deployed on the production line for testing and optimization. After a period of debugging, the system was officially launched. Before the launch, the defect miss-detection rates of the enterprise's glass and non-woven fabrics reached 15% and 12% respectively, and the false-alarm rates were 10% and 8% respectively. After the launch, the defect miss-detection rates decreased to 5% and 6% respectively, and the false-alarm rates decreased to 3% and 2% respectively.
The scarcer the defect samples are, the greater the value of learning only from good products. First, establish a normal baseline, and then use a small number of samples to supplement the criteria.
DaoAI's Solution and Products
DaoAI's solution is mainly based on the combination of few-shot learning and APDT anomaly detection. The products include high-precision image acquisition devices and advanced data-analysis software. The image acquisition devices can quickly and accurately collect images of the glass and non-woven fabric surfaces, providing data support for subsequent analysis. The data-analysis software uses few-shot learning and APDT algorithms to process and analyze the collected images and identify surface defects. This product has the characteristics of being able to go live with a small number of samples and being optimized for the semi-transparency of glass and the weak textures of non-woven fabrics, which can effectively improve the accuracy and efficiency of detection.
Quantitative results: By using DaoAI's detection solution, the enterprise achieved a defect detection rate of over 94% for the surface defects of glass and non-woven fabrics. At the same time, the defect miss-detection rates decreased by 10% and 6% respectively, and the false-alarm rates decreased by 7% and 6% respectively. The quality inspection has changed from relying on experience-based visual inspection to standardized online judgment. New defect types can be quickly supplemented with a small number of samples, and the consistency and maintainability of detection have been improved simultaneously, providing strong support for the enterprise to improve product quality and reduce production costs.
FAQ
What challenges does the surface defect detection of glass and non-woven fabrics face?
Glass is semi-transparent and is easily affected by background and reflection interference, with weak defect contrast. For example, the contrast between defect signals and background signals is less than 10%. Non - woven fabrics have disordered fibers and are weak-texture surfaces, with a texture complexity index exceeding 80, making it difficult to separate defects from the background. Moreover, the defect samples of both materials are scarce and have diverse morphologies.
How does DaoAI solve the problem of scarce defect samples?
DaoAI combines few-shot learning and APDT positive-sample anomaly detection. It first uses sufficient good-product images to establish a normal baseline, and then uses a very small number of defect samples to fine-tune the criteria. It can go live with a small number of samples, reducing the dependence on a large number of annotated samples.
What effects does the implementation of DaoAI's detection solution have?
The enterprise can quickly complete the detection system implementation even with scarce defect samples. The defect detection rates of glass and non-woven fabrics reach over 94%. The miss-detection and false-alarm rates are significantly reduced. Quality inspection has changed to standardized online judgment, and the consistency and maintainability have been improved.
This article was generated by AI. Customer cases are simulated scenarios based on real product capabilities and figures are illustrative; see product pages for official benchmarks.