---
license: apache-2.0
datasets:
- Smolry/HSRP_classification_data
language:
- en
metrics:
- name: F1
  type: f1
  value: 0.86
- name: Precision
  type: precision
  value: 0.87
- name: Recall
  type: recall
  value: 0.85
- name: Accuracy
  type: accuracy
  value: 0.87
base_model:
- google/efficientnet-b0


pipeline_tag: image-classification
tags:
- computer-vision
- license-plate
- hsrp
- traffic

library_name: pytorch

---
# HSRP Classification using EfficientNet-B0

## Model Overview

This model performs binary classification of vehicle license plate images.

Given a cropped license plate image, the model predicts whether the plate belongs to:

- HSRP
- Non-HSRP

The model forms the classification component of the Smart HSRP Detection and Violation Monitoring System.

---

## Problem Statement

Determine whether a detected vehicle license plate is an HSRP plate.

Input:
    License Plate Crop

Output:
    HSRP Probability

Decision:
    HSRP / Non-HSRP

---

## Pipeline Role

```text
Camera Frame
      ↓
Vehicle Detection
      ↓
License Plate Detection
      ↓
Plate Crop
      ↓
┌──────────────────────┐
│ EfficientNet-B0    │
│ HSRP Classifier    │
└──────────┬───────────┘
           ↓
     HSRP Probability
           ↓
     Downstream Logic

```
---

## Scope

This model performs:

✓ HSRP Classification

This model does NOT perform:

✗ Vehicle Detection
✗ License Plate Detection
✗ OCR
✗ License Plate Recognition
✗ Vehicle Tracking
✗ Violation Decision Making

---

## Model Architecture

```text
Backbone:
    EfficientNet-B0

Initialization:
    ImageNet Pretrained Weights

Classifier Head:

Dropout(0.2)
        ↓
Linear(1280 → 1)

Output:
    Binary Classification Logit

Activation:
    Sigmoid

Input Size:
    224 × 224 RGB

```
---

## Dataset

Dataset Repository:

[![HSRP-Classification Dataset](https://img.shields.io/badge/Dataset-00000)](https://huggingface.co/datasets/Smolry/HSRP_classification_data)

Classes:

0 → HSRP
1 → Non-HSRP

---

## Training Procedure

### Preprocessing

Training:

- Resize (224 × 224)
- ColorJitter
- RandomRotation (±5°)
- ImageNet Normalization

Validation / Test:

- Resize (224 × 224)
- ImageNet Normalization

---

### Phase 1 – Classifier Training

```text
EfficientNet-B0
        ↓
Frozen Backbone
        ↓
Train Classifier Only

Optimizer:
    Adam

Learning Rate:
    1e-4

Epochs:
    5

Scheduler:
    ReduceLROnPlateau

```
---

### Phase 2 – Fine Tuning

```text
Load Best Phase-1 Checkpoint
        ↓
Unfreeze Final Feature Blocks
        ↓
Fine Tune Model

Optimizer:
    Adam

Learning Rate:
    1e-5

Epochs:
    20

Scheduler:
    ReduceLROnPlateau

```
---

## Evaluation

### Test Metrics

| Metric | Score |
|----------|----------|
| Accuracy | 87.47% |
| Precision | 87.92% |
| Recall | 85.85% |
| F1 Score | 86.87% |

Test Samples:
    439

---

## Confusion Matrix

![Confusion Matrix](evaluation/confusion_matrix.png)

---

## Training History

The model was trained in two stages:

1. Classifier training with the EfficientNet-B0 backbone frozen.
2. Fine-tuning of the final feature blocks.

| Phase 1 Loss | Phase 2 Loss |
| :---: | :---: |
| ![Phase 1](training/phase1_loss.png) | ![Phase 2](training/phase2_loss.png) |

The epoch-level training history is available in
[![Training Hisitory](https://img.shields.io/badge/Training_History-training__history.json-blue?logo=json&logoColor=white)](https://huggingface.co/Smolry/HSRP-classification/blob/main/training/training_history.json)

---

## Observations

Key observations discovered during evaluation:

- Model performs strong binary separation between classes.
- Errors primarily occur in visually ambiguous samples.
- Performance depends heavily on crop quality.
- Explainability analysis should be used to validate focus on plate regions.

---

## Limitations

- Binary classification only.
- Not evaluated on every Indian state.
- Performance may degrade on heavily blurred images.
- Sensitive to plate crop quality.
- Domain shift may affect performance.

---

## Future Improvements

Potential future work:

- Larger dataset
- Additional state coverage
- CCTV-specific training data
- Hard-negative mining
- Better augmentation
- Threshold optimization
- Calibration analysis
- Lightweight deployment models

---

## Files

```text
model/
└── efficientnet_b0_finetuned.pth

evaluation/
├── metrics.json
├── classification_report.txt
└── confusion_matrix.png

training/
├── training_history.json
└── phase1_loss.png
└── phase2_loss.png

```
---

## Project Context

This model is one component of the larger Smart HSRP Detection and Violation Monitoring System.

Its sole responsibility is determining whether a cropped license plate image belongs to an HSRP or Non-HSRP category.

---

## Version History

### v1.0

- Initial EfficientNet-B0 implementation
- Two-stage training pipeline
- Binary HSRP classification

---