Related Experiment Video
Updated: Jul 10, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
A novel quantum convolutional neural network framework for quantum-enhanced classification of pixelated colour images
Chisomo Daka1, Somnath Bhattacharyya2,3
1Nano-Scale Transport Physics Laboratory, School of Physics, University of the Witwatersrand (Wits), Johannesburg, 1 Jan Smuts Ave, Braamfontein, Johannesburg, 2050, South Africa.
Scientific Reports
|March 28, 2026
Summary
The Novel Quantum Convolutional Neural Network (No-QCNN) shows promise for image classification on near-term quantum devices, outperforming classical models in low-data, multiclass scenarios. However, its performance may decrease with larger datasets due to current quantum hardware limitations.
Area of Science:
- Quantum Machine Learning (QML)
- Artificial Intelligence (AI)
- Computer Vision
Background:
- Classical Convolutional Neural Networks (CNNs) face scalability and efficiency challenges with large, complex image datasets.
- Quantum Machine Learning (QML) offers potential advantages through quantum parallelism and entanglement for feature extraction.
- Quantum Convolutional Neural Networks (QCNNs) are an emerging area within QML for image analysis.
Purpose of the Study:
- To introduce the Novel Quantum Convolutional Neural Network (No-QCNN), an end-to-end quantum model for image classification.
- To evaluate No-QCNN's performance on low-resolution colour image classification tasks, particularly in low-data regimes.
- To compare No-QCNN against classical CNNs on both multiclass and binary classification tasks.
Main Methods:
- Developed a hybrid quantum-classical No-QCNN model utilizing a variational quantum classifier (VQC) optimized with the COBYLA algorithm.
- Implemented a novel quantum feature map to encode spatial-chromatic information (RGB, position) into a hierarchical ZZFeatureMap.
- Benchmarked No-QCNN against a classical CNN on simulated quantum hardware using IBM-Qiskit for a six-class and a binary classification task.
Main Results:
- No-QCNN achieved 82.05% validation accuracy on a six-class problem with 50 images, significantly outperforming a classical CNN (40.00%).
- For a binary classification task, the classical CNN achieved 100% accuracy, while No-QCNN reached 89.7%.
- No-QCNN performance degraded with increased dataset size and training time, attributed to circuit limitations, data encoding, and NISQ noise.
Conclusions:
- No-QCNN demonstrates superior generalisation and reduced overfitting in low-data, multiclass image classification settings compared to classical CNNs.
- The model is well-suited for specific niches requiring correlation-rich feature extraction in low-data scenarios on near-term quantum devices.
- Results suggest a complementary role for No-QCNN in quantum-enhanced artificial vision and perception within the Noisy Intermediate-Scale Quantum (NISQ) era.
