Training receipt detection without labeling everything by hand.
A receipt-text localization pipeline that used Azure AI Vision to bootstrap annotations before fine-tuning YOLOv11.
Core stack
.93
detection accuracy
01 / Where it started
The problem I was trying to solve.
Thai receipts vary in layout, lighting, print quality, and text density. Drawing every training box by hand would have taken days before model training could even begin.
02 / What I built
My approach.
I used Azure AI Vision to generate structured annotations, converted them into YOLO format, reviewed the data, and fine-tuned YOLOv11 for text localization. That let me spend time on difficult examples instead of repetitive labeling.
03 / Architecture
How the pieces fit together.
Collect
Diverse receipt samples
04 / What changed
What came out of it.
- 01
50+ annotation hours saved
- 02
0.93 text detection accuracy
- 03
Custom YOLOv11 training workflow
What I learned
The most useful automation did not remove human review. It removed the repetitive first pass, so I could focus on the samples where judgment actually mattered.