An automatic target recognition (ATR) system is implemented using visual selective attention. Typical ATR system has 3 stages, which are target detection stage, clutter rejection stage, and target recognition stage. IR tank images are used for testing performance of each stage. Bottom-up selective attention is modeled by "Saliency map" proposed by Itti and multi-layer perceptron (MLP) is used to extract another features at local area. To combine bottom-up feature map with MLP feature map, 5 different methods are tried: pixel by pixel addition, pixel by pixel multiplication, using standard MLP for target/clutter pattern recognition, using MLP which has an exponetial term, using MLP which looks at broader area. Experimental result shows the last case has the best performance.
A clutter rejector is implemented using MLP. Three detection methods, which are bottom-up approach, MLP approach and combination of two maps, are compared at this stage. Final recognition stage is implemented using MLP. For three detection methods, overall automatic target detection performance is compared.
When two maps are combined using MLP, not only target detection rate but also overall performance is enhanced. Proposed method shows high recognition rate and low false alarm.