Fast Optical Image Compression Based on a Hybrid D2NN–ENN Autoencoder

The rapid growth of visual data in imaging and sensing motivates compression before full electronic processing. This work investigates a hybrid D2NN–ENN autoencoder for grayscale image compression, in which a passive diffractive optical encoder forms a compact latent intensity representation and an electronic neural-network decoder reconstructs the image. We evaluate optical precompression under different illumination modes, latent-region sizes, diffractive depths, mask parameterizations, and quantization levels. Increasing the latent region improves reconstruction quality only up to an intermediate range, after which the gain saturates. Coherent illumination provides a modest reconstruction-quality advantage over incoherent illumination. The best tested configurations use two or three diffractive layers. Full amplitude–phase modulation performs best, whereas phase-only modulation remains a practical lower-complexity alternative. Quantization at 8 and 4 bits preserves most reconstruction quality, while 2-bit precision causes visible degradation. Although JPEG and learned neural codecs achieve better rate–distortion performance, the proposed architecture is promising as an ultrafast optical precompression front end for specialized visible, near-infrared, and infrared sensing systems in which low encoding latency is more important than general-purpose storage efficiency.