SNAPPIX: Efficient-Coding–Inspired In-Sensor Compression for Edge Vision
Abstract
Energy-efficient imaging is essential for edge sensing, where energy is dominated by in-sensor data readout and wireless transmission. In-sensor compression can reduce this cost but faces challenges in hardware overhead, information loss, and task specificity. Inspired by the mammalian visual system, we present SNAPPIX, an in-sensor compression system that uses coded exposure (CE) for lightweight, sensor-compatible compression; learns a task-agnostic CE pattern by maximizing decorrelation among coded pixels based on efficient coding theory; and co-designs tile-repetitive CE patterns with Vision Transformers, augmented by reconstruction-based pre-training. Evaluating on action recognition and video reconstruction, SNAPPIX outperforms state-of-the-art video-based methods while reducing edge energy by up to 15.4x.