BOM Pattern Detection

Pattern Template

Drop image or click to browse

Engineering Drawing

Drop image or click to browse

BOM Pattern Detection

Zero-shot symbol detection for engineering drawings · No training data required

NCC DINOv2 ViT-S/14 Qwen2-VL-2B Zero-shot

Pipeline Architecture

Preprocess

Adaptive binarize · Denoise

<0.5s

→

NCC Matching

Multi-scale · Multi-angle · NMS

15–60s CPU

→

DINOv2 Verify

Cosine similarity · Derotate

2–10s GPU

→

Struct Filters

Chamfer · NMS · Gap filter

<1s

→

VLM Filter

Qwen2-VL · Optional

~0.4s/crop

Key Features

Fully zero-shot — no fine-tuning or labeled data
Multi-scale (0.7×–1.8×) & multi-angle (±10°, 90°)
Handles simple and complex symbol templates
Optional VLM stage for hard FP rejection
CPU + GPU support · Docker ready

Tech Stack

Stage 1OpenCV NCC · TM_CCOEFF_NORMED

Stage 2DINOv2 ViT-S/14 · PyTorch

Stage 3Qwen2-VL-2B · HuggingFace

BackendFastAPI · Python 3.11

DeployDocker · HuggingFace Spaces

System Specification

Đặc tả kỹ thuật chi tiết toàn bộ hệ thống: kiến trúc pipeline, thuật toán từng stage, các quyết định thiết kế, kết quả thực nghiệm và hướng cải thiện.

13 mục · Pipeline · Filters · VLM · Config · Results

Mở đặc tả hệ thống