Kharsa, R., Bouridane, A., & Amira, A. (2023). Advances in Quantum Machine Learning and Deep Learning for Image Classification: A Survey. Neurocomputing, 560, 126843.
Recent papers published between 2015 and 2022. Keywords, “Quantum, Machine Learning, Image, Classification, QCNN, VQC, Quantum Neural Network (QNN), QTN, QSVM, QKNN, and MNIST”
QML and QDL algorithms development can enable parallel tasks and reduce the time complexity of the classical algorithms.
Parameterized Quantum Circuits (PQC), also known as ansatzes, are unitary matrices that vary depending on a set of parameters.
Fundamental componets including VQC, QTN, QCNN.
1.1 Image classification
1.2 Encoding Method
1.2.1 Amplitude encoding
N pixels .
The number of the required qubits is .
Main limitation: computational complexity, circuit depth.
1.2.2 Angle encoding
Using the aprameterized rotation gates.
1 pixel → 1 angle.
, n is the number of qubits.
Meaning N pixels, then qubits are required.
1.2.3 Dense angle encoding
2 pixels → 1 angle/qubit
Example:
It reduces the required qubits to .
Experiment
1.2.4 Hybrid encoding
More advanced encoding methods, both amplitude and angle strategies to reduce the number of qubits and gates, thus reducing the circuit complexity and increasing the algorithm efficiency.
1.3 Quantum machine learning algorithms
1.3.1 Quantum support vector machine
1.3.2 Quantum K nearest neighbor
The first is the superposition state of the training set, the second is the testing sample, and the third is the ancillary qubit that determines the distance.
- IMPROVE: Image classification based on quantum K-Nearest-Neighbor algorithm
RGB image convert to HSB.
- Improved Handwritten Digit Recognition using Quantum K-Nearest Neighbor Algorithm
The quantum amplitude estimation algorithm, convert amplitude to quantum state.
Not provide experiment details or evaluations.
- Quantum K-Nearest-Neighbor Image Classification Algorithm Based on K-L Transform
Enhanced QKNN method, feature extraction from coloured and grey images.
1.4 Quantum deep learning algorithms
VQC, QTN, QCNN are hybrid classical-quantum models.
VQC focuses on training a number of PQCs that do not follow a specific classical architecture. It can be analog to Aritificial Neural Network (ANN) in classical terminology.
1.4.1 Mainly classical-Variational/parameterised quantum circuit
- Gender Recognition from Facial Images using Hybrid Classical-Quantum Neural Network
- Very deep convolutional networks for large-scale image recognition
Classical to quantum transfer learning.
Pre-trained CNN: feature extractor → pre-processed classical data → train a variational circuit.
- On Circuit-Based Hybrid Quantum Neural Networks for Remote Sensing Imagery Classification
3 conv, 1 flattened, 1 FC, 1 VQC, 1 FC, softmax. High accuracy.
- Quantum-enhanced deep neural network architecture for image scene classification
IBM platform. Quantum feature extraction phase, classical DNN for classification. Quantum layers in a hybrid model can outperform the state-of-the-art models with almost one-third of the parameters.
1.4.2 Mainly quantum
- Development of variational quantum deep neural networks for image recognition
- First model, PCA 784 to 10 pixels → angle encoding method.
- Second model, images are padded (power of 2) → amplitude encoding method.
- Third model, a hybrid encoding model.
demonstrate the ability of quantum to improve performance even wit a limited number of parameters and qubits.
However, the VQC can be enhanced with QCNN, which is more immune to the Barren Plateaus.
- Layerwise learning for quantum neural networks
- both the number of qubits and the depth of the circuit (i.e., the layers) increase to explored the Barren Plateaus Phenomenon.
- Unsuitable for NISQ hardware.
- Automatic malaria disease detection from blood cell images using the variational quantum circuit
- Minimum Redundancy Maximum Relevance (mRMR) and Principle Component Analysis (PCA) as feature redunction mehods.
- ZZ feature map data→angle.
- binary classification, comparison with classical models ~99.06% accuracy.
- Comparing Concepts of Quantum and Classical Neural Network Models for Image Classification Task
- angle embedding method: pixels multiplied by .
4.45% improvement over the classical model.
convolutional autoencoder
1.5 Quantum convolution neural network
The quantum version of the classical CNN, consists of multiple consecutive conv and Pooling layers.
Two approaches, the first is the implementation of a quantum conv filter to a classical cirucit, and the other is to implement all classification layers as a quantum circuit of quantum conv and pooling layers.
1.5.1 Quantum concolutional filter in the classical model
- Quanvolutional neural networks: powering image recognition with quantum circuits
- Quantum conv filter as the first layer.
- Simple threshold method to encoding, If the value of the image pixel is greater than the threshod, the affiliated qubit has a stats , otherwise .
- NOT applicable in actual quantum hardwware.
- Quantum Deep Learning for Steel Industry Computer Vision Quality Control.
- using variable weights
- RGB image classification with quantum convolutional ansatz
- a hybrid classical-quantum CNN that classifies RGB Images effectively.
- 12 qubits are employed to encode the classical pixels for three channels with a windows size of .
- high computation resources that are required to simulate.
- high accuracy, better performance, loss curves are smoother.
- On Depth, Robustness and Performance Using the Data Re-Uploading Single-Qubit Classifier
- a single-qubit QCNN.
- encoded using rotation gates on the same qubit.
- only six parameters.
- did not exploit the capabilities of entanglement and multi-qubit superposition.
1.5.2 Quantum convolutional network as the main model
- Design of a quantum convolutional neural network on quantum circuits
- Amplitude encoding method, overcome the ime and circuit-depth constraints, they proposed the Approximation Quantum State Preparation (AQSP).
- Quantum circuit learning
- Variational convolutional neural networks classifiers
- 变分卷积神经网络 VCNN, to replace the classical conv and pooling layers in the CNN.
- angle encoding method.
- SGD optimization. binary-cross-entropy loss function is minimised.
- Quantum convolutional neural network for classical data classification
- explored different image reduction techniques, as PCA and the autoencoder.
- disscussed three encoding methods, namely amplitude, angle and hybrid.
- the hybrid encoding method: hybrid direct encoding (HDE), hybrid angle encoding (HAE).
- small number of parameters, and high performance.
1.6 Key Challenges
1.6.1 NISQ
Most of the experiments in the discussed papers are executed in fault-tolerant simulators that do not consider the noise in real quantum computers, which questions whether the NISQ computers are capable of actually classifying images.
1.6.2 Limited number of feasible simulated qubits
The exponential growth of parameters that simulators need to keep track of with the number of qubits limits the number of feasible qubits in the simulators.
Moreover, because of this hurdle, researchers are driven to employ image reduction methods that result in immense information loss and accuracy drop.
1.6.3 Image encoding methods
Until now, all the embedding methods require image down-scaling mainly because of the considerable depth of the circuit or the number of qubits.
1.6.4 Sophisticated images datasets & Color images & Future directions
The statistics show a lack of literature supporting color and multi-class classification and indicate the necessity for new research that addresses these issues.
1.6.5 Novel image embedding
Original images without down-sizing.
1.6.6 RGB support
The red, green, and blue channels can be extracted and encoded separately using a novel approach.