pRAM nets for detection of small targets in sequences of infra-red images

A probabilistic random access memory (pRAM) neural network is described for the classification of objects in a video sequence of FLIR (forward looking infra-red) images into two classes, target and clutter. The image sequences used for training and testing were gathered from real scenes. These sequences of frames were first passed through a hot-spot detection system which identified points in the image that have a high probability of corresponding to a target. Then feature extraction was done on the image patches surrounding these hot-spots using principal component analysis (PCA). These features served as input to a reinforcement learning pRAM net trained to produce values of (1 0) for targets and (0 1) for clutter. The experimental results have been promising, and on average, the network achieved a detection probability of 0.90 and 2–3 false alarms per flame in all training and test sets.