Application issues of a programmable optical CNN implementation

A 2-D neural network and its optical implementation are presented. The neural network consists of two interconnection layers. The first layer is a modified Hamming net layer that calculates the Hamming distance between the input and each of the stored patterns. Instead of performing a winner-take-all operation, thresholding is performed within each hidden neuron. The second layer is a mapping network for pattern association. The presented neural network has fewer interconnections and a higher storage capacity than a fully interconnected Hopfield network. In comparison with multilayer perceptrons, it has advantages such as easy and direct training and unipolar binary interconnection weights and therefore is more suitable for optical implementation with currently available optoelectronic devices. The neural network can be easily reconfigured when the training set needs to be updated or extra training patterns need to be added into the training set. The attraction basins of stored patterns can be adjusted, and the input fault tolerance can be enhanced locally. An optical system employing Dammann gratings has been used in the preliminary experiments. Experimental results verified the feasibility of the neural network model as well as its optical implementation.

A new nonlinear spatial encoding method for cascadable fuzzy logic

Packet contention is a major issue in asynchronous optical packet switching networks. Optical buffering, which is implemented by fiber delay lines (FDLs), is fundamental to many optical switch implementations for resolving contention. Most existing optical buffering implementations are output-based and require a huge amount of FDLs as well as larger switch sizes, which impose extra cost on the overall system. In this paper, we consider a shared optical buffering architecture which can reduce the buffer size at a switch. We propose an analytical model to evaluate the packet loss probability and the average delay For shared buffers at a single switch. We then compare the performance of output buffers to shared buffers under different granularities of FDLs. We observe that, by choosing an appropriate granularity, the shared buffering scheme can significantly reduce packet loss with much smaller switch sizes and fewer FDLs than the output buffering architecture. The accuracy of the analytical model is also confirmed by extensive simulation

To overcome shortcomings of the currently available optical computer architectures a programmable opto-electronic CNN array computer (POAC) is introduced. It combines the optical correlators' supreme resolution and complexity with the versatility of the single chip adaptive visual CNN-UM sensor and processor arrays. The system is grounded on an unconventional optical correlator architecture, which is a modified joint Fourier transform correlator. This architecture uses the angular coding of the template pixels and applies the template not in the recording phase, but in the reconstruction phase of correlation. Feasibility of this concept has been successfully demonstrated by a breadboard model. It computes correlation in two consecutive steps: First a hologram of the input image is recorded in a dynamic holographic material (we used chemically-modified Bacteriorhodopsin film samples). Next, angularly-coded read-out beams reconstruct the appropriately weighted and shifted replicas of the stored input image. Superposition of these copies produces the correlogram. A programmable adaptive sensor array, a special Visual CNN-UM chip detects the resulting correlogram. So, both the sensing and the required post-processing steps are accomplished in a parallel way. The introduced opto-electronic architecture provides exceptional synergy and good compromise of the applicable technologies and the system performance.

We propose an optical implementation of iterative Fourier transform algorithm using a spatial light modulator and phaseshifting digital holography technique. It is shown that the optically implemented iterative Fourier transform system can improve quality of target diffraction images in real time by the same way the numerical iterative Fourier transform algorithm does in the numerical simulation. In the proposed implementation, the wave front of diffraction field is directly measured by the phase shifting digital holography technique and the phase modulation of light field is performed by a phase-type spatial light modulator. The feasibility and more general applicability of the proposed implementation of the iterative Fourier transform algorithm for digital holography techniques are discussed.

The National Optics Institute is currently carrying out a project in automatic target recognition in order to locate, recognize, and track potential targets (e.g. tanks or troop carriers) monitored in real time by an infrared camera. In this paper, we present an algorithm based on the Distance Classifier Correlation Filter (DCCF), a subclass of the Synthetic Discriminant Functions family which are often used to solve the target recognition problem. We describe the general approach of DCCF-based template matching and the algorithms used. The effect of the training set on the discrimination performance, the optical implementation, and a few practical results are also presented. For a three-class recognition, only 3 misclassifications have been reported on a data bank of more than 200 images.

Endowed with the superior computing speed and energy efficiency, optical neural networks (ONNs) have attracted ever-growing attention in recent years. Existing optical computing architectures are mainly single-channel due to the lack of advanced optical connection and interaction operators, solving simple tasks such as hand-written digit classification, saliency detection, etc. The limited computing capacity and scalability of single-channel ONNs restrict the optical implementation of advanced machine vision. Herein, we develop Monet: a multichannel optical neural network architecture for a universal multiple-input multiple-channel optical computing based on a novel projection-interference-prediction framework where the inter- and intra- channel connections are mapped to optical interference and diffraction. In our Monet, optical interference patterns are generated by projecting and interfering the multichannel inputs in a shared domain. These patterns encoding the correspondences together with feature embeddings are iteratively produced through the projection-interference process to predict the final output optically. For the first time, Monet validates that multichannel processing properties can be optically implemented with high-efficiency, enabling real-world intelligent multichannel-processing tasks solved via optical computing, including 3D/motion detections. Extensive experiments on different scenarios demonstrate the effectiveness of Monet in handling advanced machine vision tasks with comparative accuracy as the electronic counterparts yet achieving a ten-fold improvement in computing efficiency. For intelligent computing, the trends of dealing with real-world advanced tasks are irreversible. Breaking the capacity and scalability limitations of single-channel ONN and further exploring the multichannel processing potential of wave optics, we anticipate that the proposed technique will accelerate the development of more powerful optical AI as critical support for modern advanced machine vision.

In this paper we investigate a feasibility of an opto-electronic implementation of the diffusion neural network for contour detection. The diffusion neural network performs the Gaussian operation efficiently by the diffusion process. We apply this in producing the DOG (Difference of two Gaussian) functions, which can detect the intensity changes of the different spatial frequency components in an image. In the diffusion neural network each neuron has four connections with the four nearest neighbor neurons and a self-decay loop for a 2D image, and the connection weights are fixed-valued. Therefore the diffusion neural network is simpler and more efficient than LOG masking method in hardware or optical implementation. We implement the diffusion neural network opto-electronically using the point spread function of a spatial light modulator. This system is composed of a spatial light modulator, a 2D image sensor array, and a computer. The processing time of the system is very fast. Therefore the system has a potential applicability to the system that requires a real time processing of an image.

An entropy network mapped from a binary feature tree is described and compared with the Hamming net. Implementation of the entropy network includes extracting features from an input and making the decision according to these features. The feature extraction in the first layer of the entropy network is implemented by an optical method. Optical implementation with an incoherent shadow casting correlation system is proposed, and experimental demonstrations are given. The processing capability of the proposed optical system is discussed.

This paper provides an overview of the work completed to design, qualify, manufacture and integrate electrical and optical double barrier penetrators with the Electrically Trace Heated Pipe-in-Pipe (ETH-PiP) as part of the Neptune Energy Fenja Development Project. Typical subsea penetrator systems in the oil and gas industry, such as pumps, compressors and X-trees are designed to be retrievable, to enable periodic refurbishment as well as providing the option for replacement, if required. However, the ETH-PiP architecture makes retrieval of system components complicated and uneconomical. Both the electrical and optical dual barrier penetrator system designs have to comply with a set of ETH-PiP specific criteria, such as to be maintenance free over a 25 years service life, prevent water ingress to the pipeline, provide pressure containment for operational media (in an unlikely scenario where the inner pipe bursts) and guarantee minimum footprint to allow an optimum integration onto the Pipeline End Termination (PLET) structure. In addition, the electrical system has to comply with a medium voltage rating (i.e. 5.0/8.7kV) to ensure a wide range of possible ETH-PiP architectures. The optical system has to maintain insertion loss below 0.5dB and a back reflection below -45dB to comply with the stringent requirements of distributed temperature monitoring sensor system over long distances. The qualification program of the electrical dual barrier penetrator system was performed in accordance with IEC 60502-4 and SEPS-SP-1001. A tailor made sequence had to be developed for the optical system, based on guidance from SEAFOM-TSD-01, considering that the system partly falls outside the associated standard application. The electrical dual barrier penetrator system qualification sequence was developed in two phases; firstly, the electrical transition contacts in the feedthrough chamber were qualified in accordance with IEC 60502-4 and secondly, four electrical double barrier penetrator prototypes were manufactured to allow the completion of the qualification sequence defined as per SEPS-SP-1001. The optical dual barrier penetrator system qualification employed the manufacturing of three prototypes to execute the pre-defined qualification sequence. Following the individual qualification of the electrical and optical dual barrier penetrator systems, subsequent welding and full-scale assembly trials were performed to ensure that the maximum allowable temperatures within the penetrators would not be exceeded during welding to the PLET, and to proof test the assembly procedure. Electrical verification testing was also undertaken during these trials to verify that the integrity of the penetrators had been maintained during the assembly and that the PLET arrangement did not give rise to any electrical stresses that could result in excessive deterioration of the penetrators. Integration of the four electrical and two optical dual barrier penetrator systems to the project PLET was completed in Q1 2020, with the actual subsea installation of the first ETH-PiP section including the PLET in Q3 2020.

Several methods of one-way image transmission have been proposed in the past. Early methods1 used conventional holography, and the recent ones have used nonlinear optics.2 We present experimental studies of a nonlinear optical (i.e., real-time holographic) implementation of one early proposal.1 This technique consists of first recording a hologram of an aberrator and then reading it out with an image beam that is distorted through the same aberrator. The hologram is formed by the interference between a plane-wave beam transmitted through the aberrator (objectbeam) and another plane-wave beam (reference beam). Any image beam traversing the same path as the object beam of the holographic recording, including the transmission through the aberrator, can thus be diffracted from the hologram and will yield the distortion-free original image. We have demonstrated this technique in real time by using a thin photorefractive strontium barium niobate crystal (0.7 mm thick) as the holographic medium. Preliminary results showed that the processed image resolution, as limited by the combination of our optics and the crystal, was ~70μm. The constraints of this technique on the crystal thickness, severity of aberration, and field of view will be discussed.

MEMS-based optical switches

We discuss the chosen concepts, detailed design, implementation and calibration of the 85-electrode adaptive optics (AO) system of the Swedish 1-meter Solar Telescope (SST), which was installed in 2013. The AO system is unusual in that it uses a combination of a monomorph mirror with a Shack-Hartmann (SH) wavefront sensor (WFS) and a second high-resolution SH microlens array to aid the characterization, calibration, and modal control of the deformable mirror. An Intel PC workstation performs the heavy image processing associated with cross-correlations and real-time control at a 2 kHz update rate with very low latency. The computer and software continue the successful implementation since 1995 of earlier generations of correlation tracker and AO systems at SST and its predecessor, the 50-cm Swedish Vacuum Solar Telescope, by relying entirely on work-station technology and an extremely efficient algorithm for implementing cross-correlations with the large field of view of the WFS. We describe critical aspects of the design, calibrations, software, and functioning of the AO system. The exceptionally high performance is testified through the highest Strehl ratio (inferred from the measured granulation contrast) of existing meter-class solar telescopes, as demonstrated here at wavelengths shorter than 400 nm and discussed in more detail in a previous separate publication We expect that some aspects of this AO system may also be of interest outside the solar community.

In this paper, the practical tradeoffs for designing submillimeter wavelength imagers based on optomechanical systems combined with focal plane arrays (FPAs) are presented. The architecture of these systems differs for operation at short and long ranges. General formulas to derive the effective field of view of diffraction limited quasi-optical systems in these two scenarios are shown. These formulas can be used to evaluate the performance of a specific optical system implementation. As an application example, we present the design of an optomechanical system that can operate at both ranges in a modular approach. The presented implementation achieves an effective field of view, which is $\hbox{70}\%$ of the canonical one. The proposed solution consists of a linear FPA of eight active transceivers combined with a raster scan technique. The system for short-range scenario is a side-fed dual-reflector Dragonian architecture because of its good scanning performance when illuminated by an FPA. Thanks to the system's small aperture, the scanner is arranged after the primary mirror, without causing additional scan loss. The Dragonian system is then used to illuminate a confocal dual-reflector architecture to magnify its aperture, and can be used in the long-range scenario. The scanner in this case is before the main aperture and it has to be considered in the performance optimization of the optical system since it adds phase aberration loss.

In this paper we present a two-dimensional cellular hypercube architecture for image processing that combines features of the conventional hypercube and cellular logic architectures for 2-D computation cells. A unified theory of parallel binary image processing, binary image algebra (BIA), serves as a software tool for designing parallel image processing algorithms. To match the hardware to the software, we characterize the cellular processors using the same algebraic structure as BIA. The two-dimensional cellular hypercube image processor is a cellular SIMD machine with N2 cells and has a simple overall organization, low cell complexity and fast processing ability. An optical cellular hypercube implementation of BIA is proposed which offers parallel input/output and global interconnection capabilities which are difficult to do in planar VLSI technology.