QuFeX: Quantum feature extraction module for hybrid quantum-classical deep neural networks

TL;DR

QuFeX introduces a lightweight, trainable quantum feature extraction module designed for seamless integration into hybrid quantum-classical networks. By combining QCNN-inspired translational invariance with QuanNN-like local feature processing, QuFeX enables parallel, low-cost quantum filtering that mixes multiple input feature maps and preserves quantum feature information without discarding qubits. Embedding QuFeX at the bottleneck of a U-Net (forming Qu-Net) yields improved image segmentation performance on diverse datasets (FruitSeg30, PH2, ISBI-2012) while using far fewer trainable parameters than the all-classical baseline. The work demonstrates the practical viability of quantum-assisted feature extraction in real-world tasks and outlines hardware considerations and future directions for noise-aware deployment and broader quantum-enabled segmentation applications.

Abstract

We introduce Quantum Feature Extraction (QuFeX), a novel quantum machine learning module. The proposed module enables feature extraction in a reduced-dimensional space, significantly decreasing the number of parallel evaluations required in typical quantum convolutional neural network architectures. Its design allows seamless integration into deep classical neural networks, making it particularly suitable for hybrid quantum-classical models. As an application of QuFeX, we propose Qu-Net -- a hybrid architecture which integrates QuFeX at the bottleneck of a U-Net architecture. The latter is widely used for image segmentation tasks such as medical imaging and autonomous driving. Our numerical analysis indicates that the Qu-Net can achieve superior segmentation performance compared to a U-Net baseline. These results highlight the potential of QuFeX to enhance deep neural networks by leveraging hybrid computational paradigms, providing a path towards a robust framework for real-world applications requiring precise feature extraction.

Figures (16)

• Figure 1: Typical CNN architecture. The network consists of alternating convolutional and pooling layers for feature extraction, followed by fully connected layers usually for classification. The figure also illustrates the convolution operation, showing how a filter interacts with the input data to produce feature maps.
• Figure 2: Typical QCNN data processing structure. The QCNN integrates quantum circuits with a hierarchical structure analogous to classical CNNs, including quantum convolutional layers and pooling operations. The figure illustrates how quantum gates process input data to extract features, followed by measurement-based pooling to reduce dimensionality.
• Figure 3: Typical QuanNN architecture. The architecture incorporates quanvolutional layers, where classical input data is processed through quantum circuits to extract quantum-enhanced features. These features are then passed to classical layers for further processing
• Figure 4: Example of a QuFeX architecture. The QuFeX incorporates ideas and techniques from QCNN and QuanNN. In the data flow we use information taken from different feature maps (from two maps in this example) and pass it to a convolutional-inspired quantum circuit. Example of such circuit is illustrated on the top-right part of the figure where $U_1$ and $U_2$ depend on trainable parameters (denoted in the paper by $\vec{\theta}$), see Appendix \ref{['app:fruit']} for more details. Single-qubit measurements are then performed and the average of the corresponding Pauli-$Z$ operator is then passed on to produce an output feature map. In the QuFeX architecture we use different circuits (i.e., circuits on different qubits) to analyze different patches of the input data, in parallel. The circuits that run in parallel contain the same trainable set of parameters.
• Figure 5: U-Net and Qu-Net architectures. (a) The U-Net architecture consists of an encoder-decoder structure with skip connections, enabling efficient feature extraction and precise localization for image segmentation. (b) The Qu-Net architecture builds on the U-Net design by replacing the classical bottleneck layer with a QuFeX module. This hybrid architecture integrates quantum circuits for enhanced feature representation while retaining the encoder-decoder structure and skip connections for segmentation tasks.