Photonic Neuromorphic Accelerator Combined with an Event-Based Neuromorphic Camera for High-Speed Object Classification

In this work, we demonstrate the impact of an unconventional convolutional photonic accelerator, based on an optical spectrum slicing (OSS) [1], on the classification accuracy of objects, generated through a high-frame rate neuromorphic event-based camera [2]. The experimental setup is depicted in Fig. 1a. It consists of a 5 mW-632 nm-LED source, two objective lenses with NA=0.65 that concentrate the light beam into a $100\ \mu \mathrm{m}\times 100\ \mu \mathrm{m}$ channel. In our experiments, the targeted objects consist of test spheres, of different diameters (12, 16, and $20 \mu{m}$ ), used in aqua solutions. A pump regulated the sphere's speed to 0.8 m/s. The objects were recorded by a 10 kframe/s capable neuromorphic camera, with a temporal resolution of $1\ \mu \mathrm{s}$ . The camera detects pixel's contrast changes (events), similar to biological systems [2]. The recorded events were exported into 1 kframes/sec videos through a synthetic frame generator software, resulting to 2988 images per particle [2]. Images were post-processed using only noise reduction by frame subtraction and image cropping so as to reduce data volume to $100\times 100$ pixels.