When it comes to data-driven problem-solving, finding efficient, scalable solutions can make all the difference for businesses across industries. At Legentic, we’re constantly exploring new ways to leverage technology to help our clients save time and resources. This case study highlights one such solution: extracting Vehicle Identification Numbers (VINs) from large datasets of vehicle images. Read on to learn how we tackled this challenge and delivered results that truly made an impact.
A Swiss client approached us with a unique request: extract VINs from images of vehicle documents—but there was a catch. The dataset was vast, containing thousands of images, and only a small fraction of them featured the relevant documents. Applying Optical Character Recognition (OCR) to the entire dataset would have been both time-consuming and costly.
Our mission was clear: develop an efficient method to identify document images within the dataset so that OCR could be applied only where it mattered. This approach would not only reduce costs but also optimize processing time.
To get started, we needed a representative dataset that mirrored the images our client wanted analyzed. We collected images from 300 vehicle advertisements sourced from three popular Swiss platforms. We focused on ads with multiple images to avoid placeholder graphics, resulting in a dataset of 3,452 images. Among these, 125 images featured documents, such as vehicle identification forms—the key focus for our project.
Our solution revolved around a two-step workflow:
Custom Classifier Development: Using our in-house expertise, we created a classifier designed to detect document images. After rigorous testing and tuning, this model accurately identified 18,600 document images from a broader dataset of 378,103 images.
Targeted OCR Processing: We applied AWS Rekognition’s OCR technology to the document images flagged by our classifier. This step resulted in the extraction of 9,000 VINs with high accuracy.
By narrowing down the dataset to relevant images before applying OCR, we ensured a cost-effective and scalable process. The combination of our custom classifier and AWS Rekognition delivered actionable results without wasting resources.
Through this approach, we successfully:
Processed over 378,000 images.
Identified 18,600 document images.
Extracted 9,000 VINs using OCR.
This workflow highlights the power of tailored AI solutions when combined with reliable OCR technology. By focusing only on relevant images, we achieved precision and efficiency, delivering real value to our client.
Among the various methods we tested, fine-tuning a pre-trained model like MobileNet V3 Small struck the perfect balance between speed and accuracy. This fine-tuned model allowed us to detect document images with exceptional efficiency, ensuring that OCR was applied only where needed—a crucial factor in keeping costs down.
The technology is now live on the Legentic Platform, and in its first week, it successfully extracted over 2,000 VINs from vehicle documents in Switzerland.
This project was more than a technical success; it demonstrated the tangible impact of AI-driven solutions. In Switzerland, where no alternative methods exist to obtain VINs at scale, our approach proved to be both innovative and essential.
By leveraging fine-tuned AI models and OCR technology, we delivered a reliable, cost-effective solution that addressed a real-world challenge.
At Legentic, we’re committed to pushing the boundaries of what data solutions can achieve. From image classification to advanced analytics, we’re always exploring new ways to unlock insights and create value for our clients.
This project is just one example of how targeted AI applications can drive efficiency and innovation. We’re excited to continue this journey and develop more cutting-edge solutions that solve industry-specific challenges. Stay tuned for what’s next!