A target detecting neural network architecture for serial sensor data streams

Although much research has been accomplished in the use of neural networks in the computer vision world, nearly all of them apply to the application of a neural net work implementation directly to the back end of the sensor (i.e. a neural net behind a CCD array for target detection). This suggests that the neural net becomes an integrated part of the sensor during fabrication in silicon. A continuing problem in the present is there are many existing sensors (CCD, scanning detectors, etc.) that output their data in a serial pixel fashion. This paper investigates the design of a neural net architecture that can take noisy serial pixel sensor data as an input and report the detected target(s) position with respect to the sensor´s field of view (FOV). A neural network target detection architecture is presented based on the Multilayer Perceptron Neural Receiver defined by Watterson [1]. The neural network was trained to detect a Gaussian spot and allowed to detect the target through various levels of white Gaussian noise. The neural network target detection architecture´s performance in detecting a target was also investigated using row and column correlated types of noise. Through Monte Carlo simulations the performance characteristics are determined and represented as receiver operating characteristics (ROC) curves aid compared to the ROC curves of a Rayleigh Channel Receiver. The neural network target detection architecture was found to exhibit a 75% improvement in the probability of detection than the Rayleigh Channel Receiver. Finally, the real-time computing requirements for the neural network architecture are addressed and presented