Automated optical inspection solution for manufacturers.

Multi-tenant platform based on deep learning / 2+ year engagement / 6-person team

Automated optical inspection solution development

About the project

Automated Optical Inspection (AOI) solution employs cutting-edge deep learning models to deliver fast, precise, and reliable quality control for manufacturing and assembly lines. Equipped with real-time object detection and anomaly recognition, the system can identify even the smallest defects on complex, varied, or reflective surfaces, ensuring that only high-quality products make it to market.

This system continually enhances defect detection accuracy by utilizing advanced deep-learning algorithms, reducing false positives and overlooked errors. It adapts to various production environments, learning from real-time data and improving its detection capabilities through ongoing model retraining.

FRONTEND

Angular

Material UI

BACKEND

Nest

TypeORM

MySQL

ZeroMQ

CLOUD

Kubernetes

Azure

Docker

MACHINE LEARNING

YOLOv8

TensorFlow

Keras

Python

PyTorch

Product features

Role management system

Our system provides granular control, ensuring that users have precisely defined permissions for data, projects, and system functionalities.

Multi-tenancy

Access to the platform is provided to various tenants, each with their own designated space. Data is segregated to ensure the utmost security and confidentiality, eliminating the risk of any information leaks between clients.

CI/CD

Our Continuous Integration and Continuous Deployment (CI/CD) pipeline ensures efficient AI model updates, automated testing, and seamless deployment across production environments.

AI Solutions

Multi-Model Training

Model Caching

Image dataset annotation

Deep-learning

Data analysis

Our system utilizes multiple AI models to optimally balance real-time performance and detection accuracy. First, a lightweight model (e.g., MobileNet) conducts quick initial scans to identify potential defects. If additional validation is needed, the system automatically escalates the inspection to a more advanced model (e.g., ResNet), ensuring a thorough analysis without sacrificing speed.

To optimize execution time and reduce system load, we implemented a caching mechanism for the model's weights. This significantly accelerates the startup process, as the model uses pre-saved parameters without the need to reload or recompute them. Caching the weights also reduces resource consumption, which is crucial for efficient real-time operation, minimizing delays and improving overall system performance. As a result, the solution became faster and more cost-effective, especially when processing large volumes of data.

Our solution enables detailed image inspection and categorizes images into specific classes. This allows users to organize pictures into structured datasets for ongoing training and retraining of deep learning models.

The system leverages a range of neural networks, including YOLOv8 and TensorFlow 2, to enable real-time object recognition, defect detection, and anomaly segmentation. Each of these networks is re-trained on a case-by-case basis to ensure optimal performance for specific applications.

The system incorporates ChatGPT-based analysis, which goes beyond basic defect detection by providing operators with human-understandable insights. Instead of just viewing raw detection results, operators receive detailed information about the nature of defects, potential causes, and tailored recommendations for corrective actions. This enhanced analysis helps identify underlying issues and empowers operators to take proactive measures, improving system efficiency and preventing recurring problems.

How it works

Though the work is still in progress, I’m happy with the results so far. We’ve managed to release a few key features recently and I’m really enjoying the velocity we’re having with Radency. There hasn’t been anything they can’t do.