Home >>Solution>>


Share to:

Machine growth deep learning to build an automatic defect classification system

A few years ago, when the “Alpha Dog” defeated the top human Go players, people began to look forward to the day when AI will bring changes to the world. It is hoped that AI can reach or at least approach the basic recognition and judgment capabilities of human beings, thereby freeing people from complicated tasks.

In fact, with the development of the production process of LCD panels, PCB boards, semiconductor wafers and other products, the trend of product miniaturization and complexity, and the overall demand for intelligent changes in the electronics industry, intelligent machine detection technology is gradually playing Plays an increasingly important role.

Take the AOI (Automatic Optical Inspection) technology, which has a high application rate in the electronics industry, as an example, using machine vision instead of human eye detection. The suspected defective products are screened out first, and then the types and locations of the defects are judged manually. But this method still requires high labor and time costs. The crux of the problem is that the machine vision technology itself can only be tested in accordance with the given standards, and if it does not meet, it is unqualified. However, it is difficult to automatically classify specific defects and categories, and it is impossible to achieve “experience accumulation”. Therefore, whenever a new product is introduced, AOI requires complicated parameter settings and long-term personnel training to have defect detection and classification capabilities.

In response to this situation, Delta applied deep learning and artificial intelligence technology to develop an automatic defect classification system. When the machine detects the defect, it can also classify the unqualified products according to the problem and location of the defect. And in this process, continue to accumulate experience and gain growth. In turn, the defect recognition rate, defect classification accuracy rate and detection efficiency are significantly improved.

Delta’s automatic defect classification system is an intelligent visual inspection solution that combines artificial intelligence and big data analysis technology. Through machine learning technology, a system of rapid learning, accurate classification, accurate discrimination and accurate positioning of defects is built. Types are troublesome; when new products are imported, there is no need to go through cumbersome parameters. The automatic defect classification system automatically obtains the defect characteristics of the input image, and forms a detection and discrimination ability model.

Delta Automatic Defect Classification System

01 .Defect judgment and classification

Through the image training system with defect discrimination knowledge, the ADC deep learning system becomes an artificial intelligence model with defect judgment knowledge, with automatic defect classification and discrimination capabilities, and full coverage of real-time detection of various production defects. In addition, the system also corrects the error results through human-computer interaction, and feeds the correction results to the ADC in real time, allowing the artificial intelligence model to strengthen expert knowledge and improve the effect of defect discrimination.

02. Flexible information integration platform

It adopts modularization and uses the smallest necessary unit for architecture development. The functions can be replaced at will, customized functions can be easily embedded, and the image analysis algorithm can be managed through Delta’s image analysis platform to meet the needs of customers to add customized functions.

03. Training data collection tool

Provide easy-to-operate annotation tools, and users can collect training data efficiently and conveniently under the easy-to-use and smooth operation interface. In the future, it can also be used with artificial intelligence prompts to carry out highly accurate auxiliary annotations. At the same time, Delta’s data amplification technology can solve the problem of difficult data collection in actual production and reduce the training data threshold required for the model.

The automatic defect classification system consists of offline services and online services. Offline services provide smart tagging tools. After simply tagging the learning pictures, a defect model is formed based on deep learning technology to support online defect classification services. The online service analyzes the product picture to determine whether the quality is qualified; if it is not qualified, it will report the type, type and location of the defect, and transmit the test results to other systems for analysis.

Delta’s automatic defect classification system supports a variety of deployment methods such as private cloud and public cloud. It can meet the different needs of customers and apply across fields, including industries that require a large number of defect detection, a wide variety of defects and high requirements for real-time detection, such as Ceramic substrate and passive component inspection process; its core technology can also be applied to quality management of delicate industries with low tolerance to defects, such as hand-made products, and automated security monitoring, such as license plate recognition, personnel recognition, etc.

Delta’s automatic defect classification system has been introduced into many enterprise applications. For example, a panel manufacturer, due to its wide range of product types, large daily output, and frequent product switching, in order to achieve automatic defect classification and automatic identification, the panel manufacturer introduced the automatic defect classification solution independently developed by Delta, which has verified that it can bring customers Come direct and indirect economic benefits.

Direct economic benefits: The defect recognition speed has been increased from 2~3s/sheet of manual identification to 250ms/sheet, the product yield rate has increased by 5%-6%, and the missed detection rate is less than 0.5%, which can replace 60% of defect classification personnel (using panel Industry as an example); through defect type training and iteration, the defect recognition rate can reach more than 95%.

Indirect economic benefits: reduce product return rate, reduce return review costs and brand reputation loss, and optimize product technology and manufacturing process through defect type and defect location analysis.