Sri Lankan Stamp Matcher ResNet18
This model is a fine-tuned ResNet18-based stamp matcher for retrieving visually similar Sri Lankan stamp images from a known collection.
The model uses a ResNet18 backbone with a 256-dimensional embedding head. It was fine-tuned using supervised contrastive learning and hard-negative triplet fine-tuning.
Intended Use
The model is designed for stamp similarity retrieval. Given a cropped stamp image, it generates an embedding vector that can be compared with stored collection embeddings using cosine similarity.
Model Details
- Backbone: ResNet18
- Pretraining: ImageNet
- Embedding size: 256
- Training method: supervised contrastive learning
- Fine-tuning method: hard-negative triplet training
- Similarity metric: cosine similarity
Dataset
The model was trained using the Sri Lankan Stamp Matcher Dataset:
https://huggingface.co/datasets/nethsith/sri-lankan-stamp-matcher-dataset
Evaluation
One-reference collection retrieval test:
- Reference collection size: 353
- Query images: 674
- Exact top-1 accuracy: 0.875
- Exact top-5 accuracy: 0.985
- Family top-1 accuracy: 0.944
Validation split evaluation:
- Exact top-1 accuracy: 0.880
- Exact top-5 accuracy: 0.989
- Family top-1 accuracy: 0.978
Limitations
This is a retrieval model, not a strict yes/no ownership classifier. It returns the most visually similar stamps from the collection. For practical use, top-k results should be shown to the user for confirmation.
Performance may be affected by heavy cancellation marks, faded stamps, poor crops, lighting differences, and visually similar stamp designs.
License
This model is released under the MIT License.