Measurements-to-Tokens: A pure Transformer for image-free recognition at arbitrary sampling ratios

Derived from single-pixel imaging, image-free sensing enables efficient semantic interpretation directly from compressed measurements. However, existing deep learning methods are often constrained by fixed sampling ratios, and their architectural designs typically overlook the intrinsic physical modulation mechanisms of the sensing process. In this work, we introduce Measurements-to-Tokens (M2T), a unified framework leveraging a pure Transformer architecture for image-free recognition at arbitrary sampling ratios. Specifically, M2T adapts to varying sampling ratios by employing arbitrarily cropped long-range observation sequences. By treating the sensing process as a sequence modeling task, we explicitly integrate the physical correlation between intensity measurements and their corresponding modulation patterns into semantic tokens. This design enables the network to naturally process inputs of variable lengths, effectively decoupling the model architecture from specific sampling ratios. Extensive analysis demonstrates that M2T achieves state-of-the-art recognition accuracy and adapts to arbitrary sampling ratios. At a 1% sampling ratio, it reaches 96.51% average accuracy on MNIST and outperforms competing methods by 4.15 percentage points on average across two datasets, while remaining robust to noise.