ACI OS Few-Shot Self-Training: Micro-Crack Detection on Glass Surfaces

DaoAI ACI OS operating system (featuring visual foundation model for feature recognition, 5-minute zero-code automatic programming with one good sample, APDT positive/few-shot learning with 1–20 good samples, semantic false positive filtering, and 100% local private deployment via SDK/API/Docker) significantly improves product quality and production efficiency by reducing the traditional false negative rate of micro-cracks on glass surfaces in the chemical materials industry from approximately 1.5% to below 0.3% through APDT few-shot self-training. In the chemical materials sector, high-quality glass products are extensively used in construction, automotive, electronic displays, and solar energy, where surface quality directly impacts product performance and user experience. Especially micron-level surface defects, such as micro-cracks, scratches, bubbles, and inclusions, are often imperceptible to the naked eye but can lead to structural failure during subsequent processing or use. Traditional inspection methods, whether manual visual inspection or rule-based AOI, face challenges in accuracy and efficiency when dealing with such subtle and variable defects, struggling to meet increasingly stringent quality standards.

Pain Points: Why This Hurdle Is So Difficult to Overcome

In the fine processing of glass, especially for glass substrates or coated glass after ultrasonic welding, surface micro-cracks are extremely difficult to identify yet highly detrimental defects. These cracks can be caused by stress concentration, thermal shock, or impact from tiny particles. Their width is typically tens of microns or less, their length varies widely, and they exhibit different visual characteristics under different lighting angles, leading to high false positive rates and persistent false negatives for traditional inspection solutions. A mid-sized chemical materials manufacturer in the glass deep processing production line has long faced three major pain points:

Firstly, **high false negative rates and safety risks**. Production line data indicated that traditional manual visual inspection and rule-based AOI had a false negative rate of over 1.5% for micro-cracks on glass surfaces. If these undetected defective glass products entered downstream processes, they could lead to brittle fracture in final products during use, especially in applications with extremely high reliability requirements such as automotive glass and display panels, resulting in serious safety hazards and massive recall costs. Secondly, **high false positive rates and production efficiency bottlenecks**. Due to inherent glass surface textures, dust, or fluctuations in ambient lighting, traditional solutions also suffered from false positive rates as high as 8%–12%, leading to a large number of good products being misidentified as defective, requiring manual secondary review. According to statistics, the factory incurred over 5 hours of daily re-inspection labor due to false positives, severely slowing down the production rhythm and reducing overall production efficiency. Finally, **scarce defect samples and difficulty in model training**. Micro-cracks are sporadic defects, resulting in extremely rare defective samples, often only dozens or even a few images. Traditional deep learning models typically require thousands or even tens of thousands of defect samples for effective training, making it difficult to iterate and optimize defect models. When new defect types appeared, changeover cycles extended to days or even weeks, severely impacting the ability to rapidly iterate products and respond to market demands.

Technical Principles

DaoAI ACI OS operating system precisely addresses the challenge of micro-crack detection on glass surfaces through its core visual foundation model and APDT few-shot self-training technology. The system first leverages the visual foundation model's feature recognition capabilities to deeply learn and model the normal textures and microstructures of the glass surface, establishing a high-dimensional “normal” feature space. This enables the system to identify any minute anomalies deviating from normal features with extremely high precision, even micro-cracks in the tens of microns that are imperceptible to the human eye. Unlike traditional rule-based AOI, which relies on predefined thresholds and feature extractors, DaoAI ACI OS can autonomously learn from massive unlabeled data to form a robust understanding of complex glass surface textures, thereby effectively reducing false positives caused by background interference.

More critically, DaoAI ACI OS's APDT (Anomaly Pattern Discovery and Training) few-shot self-training technology is specifically designed for scenarios with scarce defect samples. When new micro-crack types or a small number of defective products appear on the production line, engineers only need to provide 1–20 good samples, and the system can quickly build a defect model through self-supervised learning and anomaly pattern discovery mechanisms. This process requires no defective samples, significantly reducing the threshold and time cost for model training. Compared to traditional deep learning solutions that require thousands of defective samples for training, APDT few-shot self-training reduces the required sample size by over 99% and enables rapid changeover and deployment within a 5-minute zero-code automatic programming framework for a single good sample. The system also features semantic false positive filtering, capable of distinguishing true defects from background noise (such as dust, water stains), further controlling the false positive rate to an extremely low level; in one case study, the false positive rate was reduced by -85%.

Typical Application Scenarios

• **Micro-Crack Detection After Glass Substrate Cutting and Chamfering**: During the cutting or chamfering of glass substrates, edge micro-cracks may occur due to stress concentration. DaoAI ACI OS can perform full-scale inspection immediately after cutting to identify edge or surface micro-cracks invisible to the naked eye, ensuring structural integrity. The challenge lies in edge reflections and varying geometries interfering with detection.
• **Surface Defect Detection on Glass Components After Ultrasonic Welding**: Ultrasonic welding technology is widely used to improve production efficiency and product quality in high-precision assembly lines, but thermal stress or mechanical vibration during the welding process can cause thermal cracks or stress cracks on the glass surface. DaoAI ACI OS can precisely capture these subtle cracks in and around the welding area, ensuring welding quality. The challenge lies in the complex textures of the welding area and potential melting marks.
• **Scratch and Inclusion Detection on Coated Glass Surfaces**: During the glass coating process, dust particles or equipment friction may cause scratches or tiny inclusions on the surface. DaoAI ACI OS can perform high-precision full inspection after coating, detecting micron-level scratches, bubbles, black spots, and other defects to ensure optical performance. The challenge lies in the reflective properties of the coating layer and distinguishing various background noises.
• **Internal Bubble and Delamination Detection in Display Panel Glass Covers**: For multi-layer composite glass structures, such as display panel glass covers, tiny bubbles or inter-layer delamination may exist internally. DaoAI ACI OS, combined with multi-spectral imaging technology, can penetrate the surface to detect internal structural defects, enhancing product reliability. The challenge lies in the imaging penetration of internal defects and depth localization.

Deployment Case Study

A mid-sized chemical materials manufacturer specializing in special glass, whose production line mainly produces ultra-thin glass substrates for high-end display panels. Before introducing DaoAI ACI OS, the factory relied heavily on traditional rule-based AOI and manual visual inspection for micro-crack detection on glass surfaces. Due to the thinness, high transparency of the product, and the extremely small size of micro-cracks, traditional inspection methods were unsatisfactory. Production line data showed that before deployment, the false negative rate for micro-cracks was over 1.5%, leading to frequent customer complaints and returns; at the same time, the false positive rate remained around 10%, requiring 3 workers to perform 5 hours of re-inspection daily, resulting in high labor costs and low efficiency. Furthermore, whenever new types of micro-cracks appeared or product changeovers occurred, model adjustment and training cycles lasted several weeks, severely restricting the speed of new product launches.

After the DaoAI team intervened, they first evaluated the production line environment and existing imaging system, then deployed the ACI OS operating system. Utilizing its APDT few-shot self-training capability, the system completed the training of existing micro-crack defect recognition models within 2 hours using only 10 good sample images. During subsequent trial runs, the system demonstrated excellent performance. After deployment, the false negative rate for micro-cracks on glass surfaces in this case was stably controlled below 0.3%, with a detection rate increased to over 99.7%, significantly reducing product quality risks. Simultaneously, thanks to DaoAI ACI OS's semantic false positive filtering function, the false positive rate was drastically reduced to about 1.5%, daily manual re-inspection time was cut to less than 1 hour, freeing up at least 2 re-inspection workers, and boosting production efficiency by 15%. Most importantly, in subsequent product iterations, DaoAI ACI OS's zero-code rapid changeover capability reduced new defect model training time from weeks to 5 minutes, greatly accelerating product launch cycles. The factory's manager stated that the introduction of DaoAI ACI OS not only solved long-standing quality issues but also positioned them as a leader in a highly competitive market.

DaoAI ACI OS's APDT few-shot self-training technology, with its unique advantages of 'zero defective samples' and 'minute-level changeover,' has completely changed our perception of rare defect detection, bringing us unprecedented production flexibility and quality assurance.

DaoAI Solutions and Products

The core solution provided by DaoAI to this chemical materials manufacturer is based on the DaoAI ACI OS operating system. This system, underpinned by a visual foundation model, deeply understands and recognizes features of complex glass surface textures, achieving high-precision, robust defect detection. In practical implementation, the DaoAI team integrated the ACI OS operating system into the customer's existing production line MES system via SDK/API interfaces, enabling real-time data exchange and inspection result feedback. For model building, engineers simply upload a small number of good sample images through the intuitive interface provided by ACI OS, and utilize the APDT few-shot self-training function to automatically generate high-precision defect detection models within minutes, all without writing any code. When the production line needs to inspect new types of glass products or new micro-crack defects, model changeover can be completed in just 5 minutes, significantly reducing downtime. DaoAI ACI OS supports 100% local private deployment, ensuring the security of critical customer production data and defect samples, keeping data within the factory and meeting stringent industry data security compliance requirements. Furthermore, DaoAI World as a unified foundation continuously learns from production line feedback, constantly enhancing the model's generalization capabilities and detection accuracy, forming a brain-eye-body closed loop for continuous optimization of intelligent manufacturing. For scenarios requiring more complex 3D morphological inspection, it can also be combined with DaoAI 2D / 3D ACI equipment, utilizing proprietary 3D cameras for 3D morphological reconstruction to detect hidden solder joints or micron-level features, further expanding inspection dimensions and accuracy.

In this case, DaoAI ACI OS helped the customer achieve significant business value: production line data showed that the false negative rate for glass surface micro-cracks was reduced from 1.5% to <0.3%, greatly mitigating product quality risks. Concurrently, the false positive rate dropped from 10% to around 1.5%, saving approximately ¥100,000 in annual manual re-inspection costs. More importantly, DaoAI ACI OS reduced the time required for new defect types or new product changeovers from weeks to 5 minutes, providing the production line with unprecedented flexibility and rapid response capabilities, thereby enhancing the customer's market competitiveness. These quantified results fully demonstrate DaoAI ACI OS's leading advantages and powerful capabilities in complex surface defect detection within the chemical materials industry.

Full solution for this scenario: the full inspection solution for ACI OS · Semiconductor Package Crack Detection

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.