Unsupervised Anomaly Detection on Textured Surfaces: Model with Good Products, Accurately Locate Unknown Defects

Material surface quality inspection is a crucial step in ensuring product quality, but traditional detection methods struggle when dealing with complex textured surfaces and diverse defects. The new solution proposed by DaoAI in WeLinkirt brings a new opportunity to the industry.

In industrial production, material surface quality inspection is a crucial process. It is directly related to the appearance, performance, and ultimate market competitiveness of products. Especially in some industries with extremely high requirements for surface quality, such as high-end manufacturing and electronic information industries, even minor surface defects can damage the product's function, thereby affecting the entire production process and the enterprise's economic benefits. Take a factory that produces functional surface materials with complex textures as an example. The surface of this kind of material has natural texture undulations, and it is widely used in the shells of high-end equipment and the panels of precision instruments. However, during the production process, various defects such as scratches, indentations, dirt, and local loss of gloss will appear on the material surface. These defects not only affect the aesthetics of the product but also may have an adverse impact on its functionality. Moreover, with the continuous application of new processes and new materials, previously unseen abnormal defects will also appear, which poses a huge challenge to surface quality inspection work.

Pain Points: Why Is It Difficult?

Traditional supervised visual inspection methods face many difficulties in material surface quality inspection. From the perspective of sample collection, the defect forms are almost infinite, and it is an almost impossible task to collect sufficient samples for each type of defect. In the case of this material factory, there are dozens of surface defects, and new abnormalities will appear in new process batches. Positive and negative samples can never be exhausted, which means that the model becomes obsolete soon after going online because it cannot adapt to newly emerging defect types, and the detection effect is greatly reduced. Statistics show that the effectiveness of the traditional model drops to 30% -40% in a short time when facing new defect types.

In the actual detection process, the problem of missed detection is also very serious. The texture background on the material surface causes a large number of weak-contrast defects to be missed. For example, some subtle scratches and local loss of gloss defects are difficult to accurately identify by traditional detection systems due to their low contrast with the normal texture. The frequent failure of the rule threshold after batch switching is also a prominent problem. There may be slight differences in texture, color, etc. between different batches of materials, which makes the originally set rule threshold no longer applicable and requires readjustment. This not only increases the time cost of detection but also leads to high rework and customer complaint costs. It is estimated that the rework cost caused by missed detection and rule threshold failure accounts for 10% -15% of the total production cost, and the customer complaint rate is as high as 5% -8%.

The root cause of these problems is that the traditional supervised visual inspection method is based on the idea of exhausting defects and tries to model and detect each known type of defect. However, in the scenario of material surface quality inspection where the defect forms are almost infinite and constantly changing, this method obviously cannot meet the actual needs. Moreover, the detection system is required to identify each type of defect, but the defects themselves cannot be predefined, which creates an irreconcilable contradiction.

Technical Principle

DaoAI's APDT positive sample anomaly detection solution in WeLinkirt adopts a new approach, shifting the detection from exhausting defects to modeling normal conditions. The solution only uses good product images for training and establishes a normal baseline for normal textures based on the DaoAI World model. Specifically, through the analysis and learning of a large number of good product images, the model can accurately capture the characteristics and laws of normal textures, forming a standard 'normal template'.

In the actual detection process, any area that deviates from this normal baseline, whether it has appeared in the training or not, will be picked out as an anomaly and located at the pixel level. This method is effective because it captures the essence of the problem. Even though the defect forms are infinite, the characteristics of normal textures are relatively stable. By establishing a normal baseline, it is possible to detect anomalies more comprehensively and accurately. Compared with traditional methods, this solution does not require collecting samples and labeling for each type of defect, which greatly shortens the deployment cycle. The deployment cycle of traditional methods is usually several weeks, while the APDT solution can compress it to several days. At the same time, since it does not rely on specific defect samples, the model also has good detection ability for previously unseen new defect types, and the normal baseline can be easily migrated to new process batches.

Typical Application Scenarios

• Scratch detection: Scratches on the material surface are one of the common defects. Due to the different widths, lengths, and depths of scratches, and in the complex texture background, the contrast between scratches and normal textures is low, making detection difficult. The APDT solution can separate the scratch area from the normal texture by establishing a normal texture baseline and accurately identify the location and range of scratches.
• Indentation detection: Indentations may be caused by external forces during the production process, and their shapes and sizes are also diverse. On the textured surface, indentations may be similar to the natural texture undulations and are easily overlooked. The solution uses pixel-level positioning technology to accurately find the indentation area and avoid missed detection.
• Dirt detection: The color, shape, and distribution of dirt are irregular, and it may be integrated with the normal texture. The APDT solution can keenly capture the abnormal signals of the dirt area through the modeling of normal textures and effectively detect dirt.
• Local loss of gloss detection: Local loss of gloss defects usually manifest as local changes in surface brightness, which are difficult to detect in the texture background. The solution can accurately identify the local loss of gloss area by analyzing the brightness characteristics of normal textures, ensuring the consistency of product surface quality.

Implementation Case

A medium-sized material factory mainly produces functional surface materials with complex textures, and its products are widely used in multiple high-end fields. Before introducing DaoAI's APDT positive sample anomaly detection solution in WeLinkirt, the factory's quality inspection team relied on manual sampling inspection and rule-based algorithms and faced serious problems of missed detection and customer complaints. Manual sampling inspection is inefficient and easily affected by subjective factors; rule-based algorithms do not work well in the face of complex texture backgrounds and diverse defect types.

During the implementation process, the factory only collected good product images for modeling without the need for defect sample labeling. The entire deployment cycle was shortened from the originally expected several weeks to several days, greatly improving the project promotion efficiency. After going online, the image-level AUROC remained stable above 99%, indicating that the model has high detection accuracy. There is no need for relabeling when switching to new batches, and the model can automatically adapt to the changes in new process batches.

By introducing DaoAI's solution in WeLinkirt, the factory has achieved a transformation from traditional quality inspection methods to intelligent detection, significantly improving product quality and production efficiency.

WeLinkirt's Solution and Product

DaoAI's APDT positive sample anomaly detection solution in WeLinkirt is a complete set of solutions. The solution mainly consists of two parts. One is the normal texture modeling module based on the DaoAI World model, which can establish an accurate normal baseline through learning from good product images. The other is the anomaly detection and positioning module, which can monitor the material surface in real-time and locate the areas deviating from the normal baseline at the pixel level. The advantage of the product lies in its innovative idea of only learning from good products, which eliminates the need for defect sample labeling and greatly reduces the cost of data collection and processing. At the same time, the solution has high portability and can quickly adapt to new process batches and new production environments. In addition, it can be connected to existing production line cameras, facilitating enterprises to upgrade without replacing existing hardware equipment.

Quantitative Results

After introducing the APDT solution, the factory has achieved significant quantitative results. The image-level AUROC has remained above 99% continuously, which means that the model has extremely high accuracy in detecting anomalies. Customer complaints caused by missed detection have significantly decreased, and the customer complaint rate has dropped from the previous 5% -8% to 1% -2%, effectively improving customer satisfaction. The quality inspection manpower has shifted from full-inspection sampling to anomaly review, and the overall detection consistency has been greatly improved, with the manpower cost reduced by 30% -40%. This shows that the solution not only improves the accuracy and efficiency of detection but also optimizes the enterprise's quality inspection process, bringing significant economic benefits to the enterprise.

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.