You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Nepali Banknote Models

A collection of models for the bankNotes-OCR pipeline — a hierarchical 2-stage banknote retrieval system for Nepalese banknotes.

Pipeline Overview

Query Image
  ├── Stage 0: ConvNeXt Tiny Binary Classifier  (banknote vs. random)
  ├── Stage 1a: YOLO Denomination Detector  (11 denominations)
  ├── Stage 1b: RFDETR + Swin Signature Classifier  (20 governors)
  ├── Stage 1c: YOLO + MobileNetV3 Year Classifier
  └── Stage 2: DinoV2-Base patch embeddings + Qdrant vector search

Model 1: ConvNeXt Tiny — Binary Classifier (`best_model_convnext_tiny_20260810_1038.pth`)

Stage 0: Determines if an image is a banknote or a random image. This is the current model used by the pipeline (since 2026-08-10).


Architecture	`convnext_tiny` (torchvision)
Classes	2: random (0), banknote (1)
Input size	224x224 RGB
Normalization	ImageNet mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]
Precision	FP32
Weight size	~334 MB
Checkpoint format	dict: `config` + `model_state_dict` (+ optimizer, epoch, best_score)

from huggingface_hub import hf_hub_download
import torch
from src.models.resnet import get_model

model_path = hf_hub_download(
    "Imbatmann/nepali-banknote-models",
    "best_model_convnext_tiny_20260810_1038.pth",
)
checkpoint = torch.load(model_path, map_location="cpu", weights_only=False)
model = get_model(model_name=checkpoint["config"]["model_name"],
                  num_classes=checkpoint["config"]["num_classes"],
                  pretrained=False)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

Training checkpoints can be recreated from the deploy package (deploy-noteclassifer.zip, deploy/models/model_zoo.py). The classifier head is nn.Sequential(Dropout(0.2), Linear(in_features, num_classes)), identical to create_convnext_tiny_model in src/models/resnet.py.

Model 2: YOLO — Denomination Detector (`best.pt`)

Stage 1a: Detects and classifies the denomination of a banknote.


Architecture	YOLO (Ultralytics)
Classes (11)	1, 2, 5, 10, 20, 25, 50, 100, 250, 500, 1000
Input size	640x640 RGB
Weight size	~113 MB

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

model_path = hf_hub_download("Imbatmann/nepali-banknote-models", "best.pt")
model = YOLO(str(model_path))
results = model("image.jpg", conf=0.5, device="cpu")

Model 3: RFDETR-Large — Signature Detector (`detector.pth`)

Stage 1b: Detects the governor signature region on a banknote.


Architecture	RFDETRLarge (rfdetr library)
Detection threshold	0.3
Weight size	~128 MB

from rfdetr import RFDETRLarge
from huggingface_hub import hf_hub_download

model_path = hf_hub_download("Imbatmann/nepali-banknote-models", "detector.pth")
detector = RFDETRLarge.from_checkpoint(model_path, device="cpu")

Model 4: Swin-Base — Signature Classifier (`classifier.pth`)

Stage 1b: Classifies a cropped signature region into one of 20 Nepali governors.


Architecture	`swin_base_patch4_window7_224` (timm)
Classes	20 Nepali governors
Input size	224x224 RGB
Normalization	ImageNet mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]
Weight size	~331 MB

Governor Classes

Bharat_Raj_Pandey, Bhekh_Bahadur_Thapa, Bijaynath_Bhattarai,
Chiranjibi_Nepal, Dipendra_Purush_Dhakal, Ganesh_Bahadur_Thapa,
Hari_Shankar_Tripathi, Himalaya_SJB_Rana, Janak_Raj_Pandey,
Kalyan_Bikram_Adhikari, Krishna_Bahadur_Manandhar, Kul_Sekhar_Sharma,
Laxmi_Nath_Gautam, Maha_Prasad_Adhikari, Narendra_Raj_Pandey,
Pradhumna_Lal_Rajbhandari, Satyendra_Pyara_Shrestha,
Tilak_Bahadur_Rawal, Yadav_Prasad_Pant, Yubaraj_Khatiwada

import timm
import torch
from huggingface_hub import hf_hub_download

model_path = hf_hub_download("Imbatmann/nepali-banknote-models", "classifier.pth")
model = timm.create_model("swin_base_patch4_window7_224", pretrained=False, num_classes=20)
model.load_state_dict(torch.load(model_path, map_location="cpu"))
model.eval()

Model 5: Year Detector + Classifier (`year/best.pt`, `year/best_classifier.pt`)

Stage 1c: Detects the year region (YOLO) and classifies the Nepali year (e.g. "BS 2077").


Detector	YOLO (`year/best.pt`)
Classifier	MobileNetV3-Small (`year/best_classifier.pt`)

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

det = hf_hub_download("Imbatmann/nepali-banknote-models", "year/best.pt")
model = YOLO(str(det))

File Inventory

File	Stage	Architecture	Precision	Status
`best_model_convnext_tiny_20260810_1038.pth`	0 — banknote	ConvNeXt-Tiny	FP32 (~334 MB)	CURRENT (2026-08-10)
`best_model.pth`	0 — banknote (legacy)	ResNet101	FP32 (~177 MB)	legacy fallback
`best_model_fp16.pth`	0 — banknote (legacy)	ResNet101	FP16 (~86 MB)	legacy fallback
`best_model_quantized.pth`	0 — banknote (legacy)	ResNet101 (quantized)	int8	legacy
`best.pt`	1a — denomination	YOLO	FP32 (~113 MB)	current
`detector.pth`	1b — signature detect	RFDETR-Large	FP32 (~128 MB)	current
`classifier.pth`	1b — signature classify	Swin-Base	FP32 (~331 MB)	current
`year/best.pt`	1c — year detect	YOLO	FP32	current
`year/best_classifier.pt`	1c — year classify	MobileNetV3-Small	FP32	current
`year_best.pt`	1c — year (legacy)	YOLO	—	legacy duplicate of `year/best.pt`

The app code (src/services/banknote_classifier_service.py) loads the ConvNeXt-Tiny file first, then falls back to the legacy ResNet101 files if the download fails. Old files are kept for rollback — do not delete.

Additional Model (not in this repo)

Model	Source	Purpose
DinoV2-Base	`facebook/dinov2-base`	Stage 2 — patch-level embeddings for visual similarity search

License

MIT

Downloads last month: -; Downloads are not tracked for this model. How to track