Two-dimensional target detection under noisy conditions with neural networks

This thesis developed a method to detect a twodimensional target in a noisy environment by using a secondorder neural network that has both translation and 90-degree rotation invariances. The target had a 5X2 constant s1ze, but could be anywhere in a 16X16 picture scene. In addition, the background contained an experimentally generated uncorrelated noise. Due to the sizes involved, the target loses its basic shape for any rotation other than 90 degrees. Therefore, only 90-degree rotation and translation invariance were employed. Backpropagation learning with the least mean square algorithm was used to train the NNW. The output function was a sigmoid. Four variations of the second-order NNW were examined. Networks with product terms only, with product and square terms, with product and linear terms, and with product, square, and linear terms were used. Only the NNW with all three terms learned and tested well under noisy conditions for the rectangular target detection. The noise levels used were 0 1 0 o 6 1 0 o 7 1 and 0 • 8 o The training scene size was limited by the s1ze of the target and chosen to be 6X6. A 6X6 moving window was used to cover the 16X16 scene. This approach reduces the size of the network and improves the convergence during training. However, the partial target problem makes the selection of threshold values for target identification critical. The NNW, trained with experimental uncorrelated noise with noise levels of 0.6, 0.7, and 0.8, had above 80% accuracy when tested with the 6X6 and 16X16 samples of corresponding levels of noise. The results have shown that the special neural network architecture can be used to detect a two-dimensional rectangular target in a large, noisy scene.