Signal, Image and Video Processing Обработка сигналов

Encoding strategies for simulated prosthetic vision based on gestalt psychology

Wang, Yuxi, Zhao, Ying, Fan, Ziyang, Gan, Lin, Han, Yuerong
Springer Nature · 2026-11-16 · DOI 10.1007/s11760-026-05773-w

Аннотация

Simulated prosthetic vision (SPV) requires selective encoding under severe spatial bandwidth constraints. This study evaluated whether four Gestalt principles—proximity, continuity, closure, and figure-ground organization—could be operationalized as stimulus-level encoding rules. Eleven conditions, including two references, four single operators, and five hybrids, were evaluated with separate computational observers on a DAVIS 2017 static classification task and a 24-class UCF101 action-recognition task. Proximity-based non-uniform sampling showed the most consistent performance pattern, while the proximity–continuity hybrid achieved the highest mean accuracy on UCF101. In the resolution analysis, proximity-based sampling exceeded the edge-enhanced baseline at all tested resolutions, with the largest gain at $$16\times 16$$ 16 × 16 (+4.36 percentage points). Group-isolated evaluation reduced accuracy for all tested strategies but preserved the relative advantage of proximity-based sampling. These findings suggest perceptual organization as a useful framework for screening low-resolution SPV encodings, while human psychophysical validation remains necessary.

Читать у издателя →