Neurocle Increases Production Efficiency of Display Manufacturer with Incredible Inspection Accuracy

Recently, the manufacturing industry has been actively utilizing AI technology for advanced quality inspection. The display field is no exception. The competitiveness of the display industry is based on quickly identifying the cause of defects and increasing yields through advanced inspection systems. Precise defect inspection is essential for 'tempered glass substrates', one of the display components. When tempered glass substrates go through the molding process to become a finished product, defects such as scratches, cracks, and black spots occur due to various factors. Defects in the glass substrate degrade the overall display performance and, in some cases, require rejecting, making inspection more challenging. In the display industry, where there is a need for precise inspection, vision inspection based on "deep learning technology" has recently gained prominence. Deep learning vision inspection has strengths in detecting atypical defects in complex environments and accurately recognizing complex patterns. AI deep learning vision inspection is presented as a breakthrough alternative to compensate for the limitations of machine vision inspection based on rule-based algorithms that require programming of complex rules or visual inspection that is time-consuming and costly.

A company that maximized the productivity and reliability of quality inspection by adopting AI deep learning vision inspection caught our attention. Company S, a display glass manufacturer, has successfully improved the production efficiency of quality inspection of substrate glass for FPDs by implementing Neurocle’s deep learning vision inspection solution.

▲Company S's existing inspection process

Company S had previously been conducting AOI and visual inspection sequentially, but this method had significant limitations. In the case of the AOI inspection machine equipped with a rule-based algorithm, it was cumbersome to set up all the equipment every time the work environment or the inspection standards changed. In addition, the inspection accuracy was poor in detecting atypical defects and precise defects. On the other hand, during visual inspection, inspection results were inconsistent depending on the mood and condition of the worker, and over a long period of time, the worker's concentration decreased, resulting in lower inspection accuracy. The problem of frequent under- and over-inspection, which reduced productivity and the reliability of defect detection, was a major obstacle for Company S, whose main customers are global top-tier companies.

Feeling the urgent need to introduce advanced vision inspection, Company S turned to Neurocle, a domestic AI deep learning vision specialist. Company S chose Neurocle for its overwhelming inspection accuracy, which has been recognized by leading conglomerates such as SK, LG, and Hyundai. Also, another major reason the company S chose Neurocle was for its active and seamless technical support, making it a highly preferred company in the machine vision industry.

The main objective of the project is to streamline the existing two inspection processes into one, maximizing production efficiency and product reliability.

To this end, Neurocle has proposed a solution based on its Auto Deep Learning Algorithm. Auto Deep Learning Algorithm is a core technology of Neurocle that automatically optimizes parameters and architecture to create "high-performance AI models". With this technology, users without coding or deep learning knowledge can easily create AI deep learning models with a few clicks.

Let's take a closer look at how Neuro-T and Neuro-R, Neurocle's deep learning vision software, are utilized in each process of vision inspection.

▲Workflow of Neurocle’s Deep Learning Vision Software

First, Neurocle consulted with the customer to define and categorize the different types of defects they wanted to detect. The types of defects were categorized into scratches, stains, marks, foreign substances, cracks, and black spots.

The next step was to train AI Deep Learning to recognize each type of defect. The inspector uploads the images of each type of defect to Neuro-T, a software for creating deep learning models. The more images of defects acquired from various angles and optical conditions, the better to create a high-performance model.

After uploading the images, inspectors implement the process called ‘Image Labeling’, which marks the defect areas within each image and labels the type of defect. Image labeling is the most labor- and time-consuming part of the entire deep learning project, and Neurocle's Auto-Labeling function greatly eases this resource burden. Auto-Labeling is a feature that automatically recommends areas for labeling based on existing labels. This allows users to quickly label many images, even if they only need to label a small number of images.

The next step is to create an AI deep learning model to automatically detect defects. Users can create a segmentation model based on the previously labeled data. Segmentation models are one of the AI deep learning model types that detects defective areas in an image on a pixel-by-pixel basis. At this stage, Neurocle's core technology, the Auto Deep Learning algorithm, makes it very easy for operators to create models. Auto Deep Learning Algorithm is a technology that automatically optimizes parameters and architecture to create high-performance AI models, allowing users without coding knowledge to create high-performance models with a few clicks.

The steps of acquiring and uploading defect images, labeling, and creating models are all performed in Neuro-T, a software for creating deep learning models.

Once the defect detection model is created, it is then applied to the high-resolution camera in real time. Using Neuro-R, a software for real-time image reading, inspectors can detect atypical defects on the display surface.

Since implementing Neurocle's deep learning technology, Company S has been able to reduce the number of inspectors by more than 70% and increase inspection speed by more than 300%. Also, they achieved an over-inspection and under-inspection rate of less than 3%, maximizing production efficiency and product reliability.

"Display glass is often defective due to various reasons such as malfunctions of manufacturing machines, microscopic defects caused by dust or dirt, and mistakes made by the manufacturer," said Hongseok Lee, CEO of Neurocle. "Recently, we have received many inquiries from manufacturers who have realized the limitations of conventional visual or rule-based inspection methods. With Neurocle's unique deep learning vision inspection solution that guarantees overwhelming accuracy in any inline process, we will help manufacturers maximize productivity and product reliability."