⚡ Quantum-Enhanced Deep Learning for Robust Night-Time License Plate Recognition
Complete architectural pipeline of HybridLPRNet_8Q, illustrating the flow from night-time image input through Zero-DCE enhancement, CNN extraction, the 8-qubit VQC bottleneck, and final Bi-LSTM/CTC decoding.
🔭 The Research Question
Most license plate recognition (LPR) systems fail after dark. This project asks: Can a quantum-enhanced neural network read license plates in night conditions more accurately than a purely classical system?
We run two systems head-to-head:
HybridLPRNet_8Q: Zero-DCE + CNN + 8-Qubit Quantum Circuit + Bi-LSTM + CTCClassicalLPRNet: Zero-DCE + CNN (larger) + Bi-LSTM + CTC
Representative night-time ALPR examples showing original image, synthetic night corruption, Zero-DCE enhancement, and predictions from both models.
📊 Performance Comparison & Convergence
The quantum model achieves competitive performance utilizing a dynamic 256-dimensional Hilbert space, requiring fewer overall parameters compared to the classical baseline.
Comparative training convergence plots for the Quantum (HybridLPRNet_8Q) and Classical (ClassicalLPRNet) models showing Validation Loss, CER, and Plate Accuracy over 100 epochs.
| Metric | ⚡ Quantum Model (HybridLPRNet_8Q) |
🔷 Classical Baseline (ClassicalLPRNet) |
|---|---|---|
| Parameters | ~1.2M | ~1.8M |
| GPU VRAM Usage | ~1.7 GiB | ~1.7 GiB |
| Best Val CER | 1.586% | 1.330% |
| Plate Accuracy | 92.0% | 92.5% |
| Training Platform | Kaggle T4 x2 (cuda:0 constraint) | Kaggle T4 x2 |
🏗️ Architecture Deep-Dive
1. Low-Light Enhancement (Zero-DCE)
Stage 1: Detailed view of the Zero-DCE (Deep Curve Estimation) light module, showing the 8-iteration curve refinement process for low-light enhancement.
2. Quantum Bottleneck (VQC)
Stage 4: Schematic of the 8-qubit Variational Quantum Circuit (VQC) featuring AngleEmbedding and two StronglyEntanglingLayers.
🔮 Quantum Interpretability
The 8-qubit register exhibits emergent specialization during training. Different qubits learn to react to specific visual features, such as numeral density or character boundaries.
Pre-circuit and post-circuit qubit activation maps across temporal slices for a representative license plate.
Bloch-sphere state trajectories for the 8 qubits, showing specialized rotational behavior.
💻 Reproduce the Research
All execution environments and training code are fully available in the notebooks/ directory.
Inference Example
import torch
# Load the Quantum Checkpoint
checkpoint = torch.load("quantum/latest.pth", map_location="cpu")
print(f"Loaded Epoch: {checkpoint.get('epoch', 'N/A')}")
print(f"Validation CER: {checkpoint.get('val_cer', 'N/A')}")
# Assuming HybridLPRNet_8Q is defined in your scope:
# model = HybridLPRNet_8Q()
# model.load_state_dict(checkpoint['model_state_dict'])
# model.eval()
👤 Author
Shanmukesh Bonala — VIT-AP University
Course: CSE4019 — Applications of AI (AoAI)
Semester: Winter 2025–26