Domain-specific Large Vision Models (LVMs) represent a critical innovation in artificial intelligence, providing tailored solutions to meet the distinct needs of various industries. Utilizing deep learning to process and interpret extensive visual data, these models offer insights that can significantly enhance operations, decision-making, and unlock new possibilities. Focused on specific fields like healthcare or manufacturing, LVMs surpass general-purpose models by learning from large, specific datasets, detecting complex visual patterns that broader models may miss.
Solution
Lanner collaborates with Landing AI, a leading cloud-based computer vision provider, to develop an end-to-end solution that integrates Lanner's edge AI appliance with LandingLens’ computer vision platform. The joint solution can enable Domain-Specific Large Vision Models, accurately capturing essential image features, enabling faster development times for downstream vision tasks and achieving higher accuracy with less effort in data labeling.
Use Case
One of the popular use cases for domain-specific large vision models is to enable automated defect classification (ADC) in the semiconductor industry. This system automatically categorizes anomalies on wafers into more than 15 defect classes.
Significant achievements from this solution include an enhanced classification rate, with over 90% alignment to manual human classifications, demonstrating the system's ability to match human accuracy. There's also a notable 60% reduction in labor, highlighting the efficiency of automation. Improved Quality Assurance (QA) insights have led to better process control, essential for maintaining production quality.
Source: Landing AI
Additionally, the system's scalability allows for expansion to meet growing operational needs without sacrificing performance, ensuring the inspection process remains robust as demands increase. The success of this initiative is largely attributed to the utilization of high-quality images and the expertise of subject matter experts, ensuring accurate and detailed analyses.
Additionally, early deployment and the establishment of clearly defined categories have streamlined the process, allowing for more efficient identification and classification of data.
Lanner Edge AI Appliances
Lanner's Edge AI appliances, the LEC-2290E and EAI-I131, can enable end-to-end, data-centric visual inspection solutions based on domain-specific large vision models.
Equipped with an NVIDIA® A2 Tensor Core GPU, the NVIDIA-Certified LEC-2290E rugged edge AI appliance enables AI-based video and data analytics across industrial edge applications, including healthcare, transportation, and smart cities. With its powerful NVIDIA A2 GPU, comprehensive industrial I/O, and integrated TPM security, the Lanner LEC-2290E is designed for accelerated workloads at the edge, providing customers access to advanced software capabilities for enhanced operational efficiency.
The EAI-I131, an industrial-grade AI inference appliance, harnesses the power of the latest NVIDIA® Jetson Orin™ NX or Jetson Orin Nano™ system-on-modules, making it ideal for computer vision and video analytics. This IP40-rated, fanless device boasts a wide operating temperature range from -40° C to 75° C, ensuring reliability and durability in tough industrial environments. Additionally, the EAI-I131 supports LTE, 5G Sub6, and Wi-Fi wireless connectivity, alongside a wealth of connection options such as 2x GbE PoE, 2x COM, 2x USB, and 4x DI/DO ports, catering to diverse operational needs.
Conclusion
Domain-specific large vision models lead the way in applying AI to industry-specific challenges, enhancing precision, efficiency, and innovation potential. This Computer Vision Solution, powered by deep learning, streamlines visual inspection and development for quality control in manufacturing and industrial settings. By using proprietary datasets designed for specific applications like medical imaging or semiconductor defect identification, these models efficiently detect crucial image details, speeding up vision task development and increasing accuracy with minimal data labeling effort.