Gerhard Bader

Fourier optical realization of cellular neural networks

Cellular neural networks (CNNs) consist of analog, nonlinear, dynamic processing elements which are locally interconnected. Most applications in the areas of image processing, pattern recognition and robot control require interconnections between all elements which are space invariant. This suggests an optical CNN implementation because optical processors are perfectly suited for both space invariant signal processing and complete interconnections between all elements. The theoretical and practical aspects of a hardware realization are described. The results of an optical CNN performing feature extraction are presented. >

Fast and accurate techniques for measuring the complex transmittance of liquid crystal light valves

A highly accurate measurement technique has been developed and applied for characterization of the complex transmittance of liquid crystal light valves (LC-LV). The measurement setup is based on a two-beam interference with partially coherent light. An arc lamp light source enables measurements over a large range of wavelengths. We propose a new Fourier transform interference pattern evaluation technique with a high signal-to-noise ratio. Calculations are sped up by FFT algorithms. Measurement results of the complex transmittance are shown for a twisted nematic and a Freederickzs liquid crystal light valve, built in our laboratory.

Shift-, rotation-, and scale-invariant shape recognition system using an optical Hough transform

We present a hybrid shape recognition system with an optical Hough transform processor. The features of the Hough space offer a separate cancellation of distortions caused by translations and rotations. Scale invariance is also provided by suitable normalization. The proposed system extends the capabilities of Hough transform based detection from only straight lines to areas bounded by edges. A very compact optical design is achieved by a microlens array processor accepting incoherent light as direct optical input and realizing the computationally expensive connections massively parallel. Our newly developed algorithm extracts rotation and translation invariant normalized patterns of bright spots on a 2D grid. A neural network classifier maps the 2D features via a nonlinear hidden layer onto the classification output vector. We propose initialization of the connection weights according to regions of activity specifically assigned to each neuron in the hidden layer using a competitive network. The presented system is designed for industry inspection applications. Presently we have demonstrated detection of six different machined parts in real-time. Our method yields very promising detection results of more than 96% correctly classified parts.

17.2: A Backlight System Providing Variable Viewing Angles for Transmissive LCDs

A simple optical system is described which allows for the variation of the viewing angle of transmissive-type LCDs. It is well suited for automatic teller machines (ATMs) to switch between a wide viewing angle state which is defined by the used LCD and a restricted viewing angle for privacy, where a contrast of 30 at normal viewing direction decreases to a value of less than 0.3 at an angle of 30°.

Microlens array processor with programmable weight mask and direct optical input

We present an optical feature extraction system with a microlens array processor. The system is suitable for online implementation of a variety of transforms such as the Walsh transform and DCT. Operating with incoherent light, our processor accepts direct optical input. Employing a sandwich- like architecture, we obtain a very compact design of the optical system. The key elements of the microlens array processor are a square array of 15 X 15 spherical microlenses on acrylic substrate and a spatial light modulator as transmissive mask. The light distribution behind the mask is imaged onto the pixels of a customized a-Si image sensor with adjustable gain. We obtain one output sample for each microlens image and its corresponding weight mask area as summation of the transmitted intensity within one sensor pixel. The resulting architecture is very compact and robust like a conventional camera lens while incorporating a high degree of parallelism. We successfully demonstrate a Walsh transform into the spatial frequency domain as well as the implementation of a discrete cosine transform with digitized gray values. We provide results showing the transformation performance for both synthetic image patterns and images of natural texture samples. The extracted frequency features are suitable for neural classification of the input image. Other transforms and correlations can be implemented in real-time allowing adaptive optical signal processing.

1.3-in active matrix liquid crystal spatial light modulator with 508-dpi resolution

Optical signal processing systems frequently use Fourier transform techniques for correlation, filtering, etc. In practical applications the complex-valued filters placed in the frequency domain plane are often realized by a liquid crystal spatial light modulator (LCSLM). The requirements for such LCSLMs are very demanding, as they should offer a high resolution, high uniformity, low fluctuation in time and an increased number of grayscales or phase steps. We designed and manufactured an active matrix LCSLM (AMLCSLM) and the appropriate driving system meeting these requirements. The AMLCSLM consists of 480 by 480 pixels, each containing an amorphous silicon thin film transistor (aSi-TFT). Although the pixel size is only 50 by 50 micrometer squared, corresponding to a resolution of 508 dpi, an optical aperture of more than 40 percent is achieved. By the use of a transparent storage capacitor the pixel capacitance was greatly increased, resulting in a reduction of the fluctuation in time. The driving circuitry is capable of driving up to 256 grayscales at a frame rate of 100 Hz in dual scan driving mode. For the row scanning the commonly used rectangular impulse shape was substituted by a shortened trapezoidal impulse. As a result a very good uniformity over the area with an overall phase change error of less than 6 percent is obtained.