Precision Financial Detection
Th-Slip-OCR-K is a specialized computer vision model designed as a pre-processing engine for OCR pipelines. It isolates specific regions of interest within financial documents, eliminating noise and enabling higher accuracy for text extraction systems.
This model is engineered to extract full telemetry from Kasikorn Bank (K-Bank) transaction slips, optimized for edge devices and mobile deployment.
System Architecture
The model utilizes the Ultralytics YOLOv8 Nano (v8.3.0) architecture, trained on Apple M4 Silicon using Metal Performance Shaders (MPS) for accelerated convergence.
- Architecture: YOLOv8 Nano
- Precision: 99.8%
- mAP50: 0.995
- Recall: 1.0 (Zero miss rate)
Detection Zones (Classes)
The model scans and identifies 6 critical data points for financial reconciliation:
| Label | Description | Context |
|---|---|---|
amount | Transaction Value | Net transfer amount (e.g., 100.00) |
accnum | Account Number | Detects masked formats (e.g., xxx-2-x8574-x) |
date | Timestamp | Date and Time of transaction |
pp1 | Sender Identity | Name of the person initiating the transfer |
pp2 | Receiver Identity | Name of the beneficiary |
transaction | Status | Operation status (e.g., Transfer Success) |
Implementation
Seamlessly integrate into your Python workflow using the ultralytics library. The model can be loaded directly from the Hugging Face Hub.
from ultralytics import YOLO
# 1. Initialize the model (Auto-download from HF)
model = YOLO("https://huggingface.co/Pondet/Th-Slip-OCR-K/resolve/main/best.pt")
# 2. Run Inference on a slip image
results = model("path/to/slip.jpg")
# 3. Visualize telemetry
results[0].show()Operational Scope
Domain Specific: Optimized exclusively for Kasikorn Bank (Green Theme) slip layouts.
Privacy Aware: The model accurately detects masked account numbers but does not attempt to reconstruct hidden digits without external database verification.
Use Case: Developed as a Proof-of-Concept for automated household accounting and slip verification systems.