With the advancement of deep learning and neural networks, the computational demands for applications in wearable devices have grown exponentially. However, wearable devices also have strict requirements for long battery life, low power consumption, and compact size. In this work, we propose a scalable optoelectronic computing system based on an integrated optical convolution acceleration core. This system enables high-precision computation at the speed of light, achieving 7-bit accuracy while maintaining extremely low power consumption. It also demonstrates peak throughput of 3.2 TOPS (tera operations per second) in parallel processing. We have successfully demonstrated image convolution and the typical application of an interactive first-person perspective gesture recognition application based on depth information. The system achieves a comparable recognition accuracy to traditional electronic computation in all blind tests.

Metal parts with highly dynamic areas often appear in industrial production measurements. However, if the traditional fringe projection technique is used to project fringe onto the surface of these metal parts, the light energy will be excessively concentrated and the image will be saturated, resulting thus in the loss of fringe information. To effectively address the high reflectivity problem of the object under test in fringe projection, background normalized Fourier transform contouring was combined with adaptive fringe projection in this work and a new method for performing highly dynamic 3D measurements was proposed. To reduce the number of the acquired images by the camera, a monochromatic fringe of different frequencies was put into the RGB channel to make color composite fringe, and then a color camera was used to acquire the deformed color composite fringe map. The images acquired by the color camera were then separated into three channels to obtain three deformed stripe maps. The crosstalk was also removed from these three images, and the 3D shape of the object was reconstructed by carrying out Fourier transform contouring with background normalization. From our experiments, it was demonstrated that the root mean square error of the proposed method can reach 0.191 mm, whereas, unlike the traditional methods, the developed method requires four images.

Multimode waveguide bend is one of the key components for realizing high-density mode-division multiplexing systems on chip. However, the reported multimode waveguide bends are either large, bandwidth-limited or fabrication-complicated, which hinders their applications in future high-density multimode photonic circuits. Here we propose a compact multimode waveguide bend supporting four TE modes simply by shape-optimizing with transformation optics. The shape of the waveguide is optimized in the virtual space with gradient distribution of the refractive index, so that the scattering loss and intermode cross talk are well suppressed. After conformal mapping back into the physical space, a compact (effective radius of 17 μm) multimode bending waveguide is obtained. Simulations show that the proposed multimode waveguide bend has little loss (<0.1dB) and low cross talk (<-20dB) throughout an ultrabroad wavelength range of 1.16–1.66 μm. We also fabricated the shape-optimized multimode bending waveguide on a silicon-on-insulator wafer. At 1550 nm wavelength, the measured excess losses for the four lowest-order TE modes are less than 0.6 dB, and the intermode cross talks are all below -17dB. Our study paves the way for realizing high-density and large-scale multimode integrated optical circuits for optical interconnect.

Photonic-assisted microwave frequency identification with distinct features, including wide frequency coverage and fast tunability, has been conceived as a key technique for applications such as cognitive radio and dynamic spectrum access. The implementations based on compact integrated photonic chips have exhibited distinct advantages in footprint miniaturization, light weight, and low power consumption, in stark contrast with discrete optical-fiber-based realization. However, reported chip-based instantaneous frequency measurements can only operate at a single-tone input, which stringently limits their practical applications that require wideband identification capability in modern RF and microwave applications. In this article, we demonstrate, for the first time, a wideband, adaptive microwave frequency identification solution based on a silicon photonic integrated chip, enabling the identification of different types of microwave signals from 1 to 30 GHz, including single-frequency, multiple-frequency, chirped-frequency, and frequency-hopping microwave signals, and even their combinations. The key component is a high Q-factor scanning filter based on a silicon microring resonator, which is used to implement frequency-to-time mapping. This demonstration opens the door to a monolithic silicon platform that makes possible a wideband, adaptive, and high-speed signal identification subsystem with a high resolution and a low size, weight, and power (SWaP) for mobile and avionic applications.