About ScanTire AI
Automated tire inspection through computer vision and deep learning
Our Mission
ScanTire is an automated hardware-software complex designed for non-destructive inspection and physical quality assessment of pneumatic tires. We solve the problem of automated incoming inspection, cataloging, and condition assessment for auto repair shops, warehouses, and car exchange e-commerce platforms.
The Challenge
Black-on-Black OCR Problem
Decoding information from a tire is an extremely challenging task. This is a "black-on-black" text recognition problem. The text has no color contrast, relies entirely on barely-visible physical relief, and is heavily distorted by rubber micro-texture, dirt, and uneven lighting.
Our Technology
ResNet18 Classifier
Deep residual neural network (transfer learning via ImageNet) analyzes macroscopic tread patterns and microscopic rubber fatigue for 5-class wear classification.
YOLOX Detector
Anchor-free object detection analyzes full 4K images and instantly locates the precise bounding box of target text (tire size) or brand logos.
ONNX Runtime
Cross-platform inference engine enables deployment on various platforms without dependency on specific ML frameworks. Optimized for production.
C++ Math Core
High-performance computer vision engine for preprocessing: median filtering, vignette correction, CLAHE, Sobel gradients, and adaptive binarization.
Key Capabilities
Curved OCR
Extract embossed tire dimensions (e.g., 205/55 R16)
Classification
Summer, Winter Studded, Friction, All-Season
TWI Analysis
Contactless tread depth measurement via TWI markers
AI Wear Rating
5-class classification from "New" to "Critical"
Current Status
- • Current accuracy: ~60-75% (improving with user feedback)
- • Model: ResNet18 trained on proprietary tire dataset
- • Processing: Mock mode (real ONNX inference coming soon)
- • Goal: Achieve 85%+ accuracy before Pro tier launch