In-Sensor Compressed Imaging with Reconstruction-Free Recognition via Ferroelectric Photodiodes

The transition from human-centric to machine-centric vision systems demands innovative imaging architectures. Compressed imaging, using key mathematical transformations to capture essential visual information during sampling, is suited for sensor-level integration to minimize data redundancy. Here, we propose an efficient machine vision strategy that connects in-sensor compressed imaging with neural networks. By manipulating ferroelectric polarization, BiFeO3 photodiodes exhibit 113 to 274 nonvolatile photoresponse states with good stability, linearity, and an 85% yield across 150 devices. The Hadamard product between the devices and image matrices transforms the input images from the spatial to the Haar wavelet domain, enabling in-sensor compression by discarding high-frequency coefficients. The compressed data are directly fed into neural networks, bypassing image reconstruction, and achieve a simulated accuracy of 87.0% at 0.04 compression ratio and 95.8% at 0.9. This approach merges image sensing and compression within the ferroelectric sensors, integrating reconstruction-free deep processing for an efficient machine vision solution.