SynergiProtoNet_Bangla_OCR_Few_Shot_Learning

This project implements the SynergiProtoNet, a few-shot learning model for recognizing handwritten characters and digits in Bangla script. Based on the methodologies described in our paper, this model demonstrates the ability to perform high-accuracy recognition with limited labeled data, addressing challenges inherent to low-resource languages. It also demonstrates comparative analysis among different classical few shot learning approaches.

Prerequisites

Before you run this notebook, ensure that you have the following:

• Python 3.10 or later
• Pip package manager
• Git CLI
• Access to a GPU (recommended for faster computation)

Download Datasets

• CMATERdb- https://www.kaggle.com/datasets/mostafiz53/basicfinal
• NumtaDB- https://www.kaggle.com/datasets/BengaliAI/numta
• BanglaLekha-Isolated- https://www.sciencedirect.com/science/article/pii/S2352340917301117
• Devanagari- https://www.kaggle.com/datasets/suvooo/hindi-character-recognition

Performance Analysis

Monolingual Intra-Dataset Evaluation

Network	1-shot Acc(%)	1-shot F1	1-shot Recall	1-shot Precision	5-shot Acc(%)	5-shot F1	5-shot Recall	5-shot Precision	10-shot Acc(%)	10-shot F1	10-shot Recall	10-shot Precision
Matching	69.64	0.69	0.69	0.69	38.66	0.39	0.39	0.39	36.36	0.36	0.36	0.36
Simpleshot	69.42	0.70	0.71	0.70	75.68	0.78	0.78	0.77	82.9	0.82	0.82	0.82
Relation	77.10	0.76	0.76	0.76	85.58	0.87	0.87	0.87	83.2	0.83	0.83	0.83
BD-CSPN	69.68	0.71	0.71	0.71	76.58	0.76	0.76	0.76	83.48	0.82	0.82	0.82
Prototypical	74.48	0.74	0.75	0.74	87.88	0.87	0.87	0.87	87.56	0.87	0.87	0.87
SynergiProtoNet	79.1	0.79	0.79	0.79	88.95	0.88	0.88	0.88	90.04	0.90	0.90	0.90

Monolingual Inter-Dataset Evaluation

Network	1-shot Acc(%)	1-shot F1	1-shot Recall	1-shot Precision	5-shot Acc(%)	5-shot F1	5-shot Recall	5-shot Precision	10-shot Acc(%)	10-shot F1	10-shot Recall	10-shot Precision
Matching	48.6	0.52	0.51	0.52	41.8	0.39	0.4	0.39	26.9	0.26	0.26	0.26
Simpleshot	51.34	0.55	0.54	0.54	63.54	0.63	0.63	0.63	64.94	0.66	0.66	0.66
BD-CSPN	50.84	0.5	0.5	0.51	60.3	0.6	0.6	0.6	69.2	0.7	0.7	0.7
Prototypical	54.28	0.55	0.56	0.55	76.18	0.75	0.75	0.75	76.24	0.75	0.75	0.75
Synergi w/o pretrain	44.5	0.45	0.44	0.44	65.16	0.66	0.66	0.66	68.76	0.68	0.68	0.68
SynergiProtoNet	59.02	0.59	0.59	0.59	77.68	0.78	0.78	0.78	81.36	0.83	0.83	0.83
Relation	56.46	0.58	0.58	0.58	74.02	0.72	0.72	0.72	65.04	0.67	0.65	0.65

Cross-Lingual Performance Analysis

Network	1-shot Acc(%)	1-shot F1	1-shot Recall	1-shot Precision	5-shot Acc(%)	5-shot F1	5-shot Recall	5-shot Precision	10-shot Acc(%)	10-shot F1	10-shot Recall	10-shot Precision
Matching	52.84	0.52	0.52	0.52	26.68	0.27	0.27	0.27	37.58	0.38	0.38	0.38
Simpleshot	42.92	0.46	0.46	0.46	56.14	0.53	0.53	0.53	55.10	0.54	0.54	0.54
BD-CSPN	44.58	0.45	0.45	0.45	54.72	0.52	0.52	0.53	57.72	0.57	0.57	0.57
Prototypical	53.64	0.55	0.56	0.55	76.74	0.77	0.77	0.77	79.48	0.79	0.79	0.79
Synergi w/o pretrain	50.46	0.51	0.51	0.51	71.14	0.70	0.70	0.70	76.98	0.76	0.76	0.76
SynergiProtoNet	58.59	0.55	0.55	0.55	76.84	0.77	0.77	0.77	82.12	0.82	0.82	0.82
Relation	61.12	0.61	0.61	0.61	74.02	0.72	0.72	0.72	81.93	0.82	0.82	0.82

Split Digit Testing

Network	1-shot Acc(%)	1-shot F1	1-shot Recall	1-shot Precision	5-shot Acc(%)	5-shot F1	5-shot Recall	5-shot Precision	10-shot Acc(%)	10-shot F1	10-shot Recall	10-shot Precision
Matching	73.67	0.74	0.75	0.75	33.1	0.29	0.34	0.34	32.17	0.33	0.34	0.34
BD-CSPN	55.43	0.59	0.59	0.59	64.43	0.64	0.64	0.64	69.17	0.68	0.68	0.68
Simpleshot	65.6	0.61	0.62	0.61	72.9	0.7	0.71	0.73	76.97	0.76	0.76	0.76
Relation	64.97	0.63	0.64	0.64	79.83	0.8	0.79	0.8	81.37	0.83	0.83	0.83
Prototypical	66.57	0.65	0.65	0.65	76.33	0.75	0.74	0.74	87.1	0.86	0.86	0.86
Synergi w/o pretrain	46.47	0.48	0.47	0.47	52.6	0.53	0.53	0.53	57.6	0.57	0.57	0.57
SynergiProtoNet	37.4	0.37	0.37	0.37	83.73	0.83	0.83	0.83	88.3	0.87	0.87	0.87

License

This project is licensed under the MIT License - see the LICENSE file for details.

Cite

@inproceedings{MehediAhamed,
author = {Ahamed, Mehedi and Kabir, Radib and Dipto, Tawsif and Mushabbir, Mueeze and Ahmed, Sabbir and Kabir, Md},
year = {2024},
month = {12},
pages = {1-6},
title = {Performance Analysis of Few-Shot Learning Approaches for Bangla Handwritten Character and Digit Recognition},
doi = {10.1109/STI64222.2024.10951048}
}