Speakers
Description
Semantic segmentation underpins scene understanding, yet pixel-level annotation remains the main issue in domains such as medical imaging or remote sensing. Self-supervised learning (SSL) extracts rich representations without labels, and quantum computing offers high-dimensional Hilbert-space encodings; combining the two, however, is constrained by the limited qubit budgets of the Noisy Intermediate-Scale Quantum (NISQ) era.
We introduce Q4S, a hybrid architecture integrating a 9-qubit variational quantum segmentation head with a frozen, SSL-pretrained classical backbone. The central challenge lies in the quantum-classical information bottleneck, where a limited qubit budget cannot natively ingest a full, high-dimensional feature map. We circumvent this limitation algorithmically by partitioning the backbone's ResNet output into 49 independent discrete spatial regions processed one at a time.This parallelization allows our 9-qubit circuit to process a total of $49 \times 512$ classical values across the image, decoupling the hardware qubit requirement from the spatial resolution. This process is driven by our modified STEGO contrastive loss, which enforces global semantic consistency across the independent regions. By extracting the variational circuit's Pauli-Z expectation values, we obtain bounded quantum outputs within the $[-1, 1]$ interval. This inherently matches the bounded nature required for cosine similarity computations, enabling seamless integration with the classical loss framework and allowing a fixed NISQ device to handle feature maps of arbitrary size.
Evaluated on a 100-image subset of the COCO-Stuff dataset, this image-by-image pipeline achieves a Pixel Accuracy of 86.29% and an MPA of 79.71%, complemented by an F1 score of 0.6445 and a mIoU of 0.5526.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|