Vision Transformer for Industrial Defect Detection

A Swin Transformer model identifying manufacturing anomalies in real-time.

Source

99.1%

F1-Score

45 FPS

Inference speed

22ms

Orin Latency

The problem

Legacy computer vision systems fail under varying factory lighting conditions and miss tiny surface cracks on metal parts.

The approach

We trained a Swin Transformer (Swin-B) model with self-supervised pre-training, fine-tuned on custom industrial scans. Quantized to INT8 and compiled using NVIDIA TensorRT for edge inference.

Results

Detected sub-millimeter defects with 99.1% F1-score at 45 FPS on an NVIDIA Jetson Orin Nano, outperforming legacy systems by 15%.