A spectral target recognition technique has been developed that detects targets in hyperspectral images in real time. The
technique is based on the configuration of a radial basis neural network filter that is specific for a particular target
spectral signature or series of target spectral signatures. Detection of targets in actual 36-band CASI, and 210-band
HYDICE images is compared to existing recognition techniques and results in considerable reduction in overall image
processing time and greater accuracy than existing spectral processing algorithms.
An object recognition technique has been developed that allows the rapid screening of multispectral images for objects with known spectral signatures. The technique is based on the configuration of a radial basis neural network (RBN) that is specific for a particular object spectral signature or series of object spectral signatures. The method has been used to identify features in CASI-2 and HYDICE images with results comparable to conventional spatial object recognition techniques with a significant reduction in processing time. Radial basis neural networks have several advantages over the more common backpropagation neural networks, including better selectivity and faster training, resulting in a significant reduction in overall image processing time and greater accuracy.
Hyperspectral image processing techniques are utilized for a variety of applications from geological surveys to detection of camouflaged enemy vehicles. One of the persistent problems is that huge amounts of data must be processed, since a hundred or more frequency bands of spectral information can make up a typical hyperspectral image cube. If real time processing is necessary, as in target tracking or identification, some means of selecting which bands are relevant to the image and which bands can be safely ignored is desirable. We propose a fast, easily trainable neural network filter architecture that can rapidly screen a hyperspectral image cube in near real time. A bank of filters, operating in parallel, is used to screen an image for suspected targets. Performance on simulated and real images is compared to existing recognition techniques and results in considerable reduction in overall image processing time and greater accuracy.
This paper describes the separation of merged signals from a mass-selective chromatographic detector by means of an adaptive filtering technique. The technique is based on parallel feed-forward neural networks, which are trained to resolve the mass spectra of two merged chemical compounds. Specifically, the chemical mass spectra of the compounds ethyl benzene and xylene were used to evaluate a filter based on probabilistic neural networks (PNN). The results are that the PNN filter shows good noise rejection and is fast enough computationally to be utilized in real time. The filter technique has applications in on-line processing of environmental monitoring instrumentation data and direct processing of pixel spectral data, such as hyperspectral image cubes.