Using Back Propagation Algorithm and Genetic Algorithm to Train and Refine Neural Networks for Object Detection

We describe a two-stage approach to the use of pixel based neural networks for object detection problems in which the locations of relatively small objects in large pictures must be found. The networks use a squared input eld which is large enough to contain all objects of interest. In the rst stage the network is trained on examples which have been cut out from the large pictures. A back error propagation algorithm and a genetic algorithm are used for the network training. The tness function for this genetic algorithm is based on the total sum squared error of the cut-outs (sub-images). The two trained networks are then applied, in moving window fashion, over the large pictures to locate the objects of interest. In the second stage the weights of the two trained networks are adjusted and rened using a second genetic algorithm. The tness function in this case is based on the precision and recall performance of the network on the full training images. The two algorithms in the rst stage can be used as two independent methods for object detection on full test images. Di erent combinations of the three algorithms in the two stages produce two new detection methods. The goal of the approach is to investigate whether the rst genetic algorithm can have better test performance on the sub-images than the back error propagation algorithm in stage one, to test whether the second genetic algorithm can improve the performance of object detection on the full test images in the second stage, and to nd which of the four methods has the best results for object detection across the two stages. We have tested all the methods on three object detection problems of increasing di culty. Regarding training the large networks on the image data, the results show that the back error propagation algorithm has stronger generalisation ability than the rst genetic algorithm. In all cases the second genetic algorithm resulted in improved deProceedings of the 1998 Computer Science Postgraduate Students Conference, Royal Melbourne Institute of Technology, Melbourne, Australia, December 8, 1998. tection performance over the networks trained by either the backward error propagation algorithm or the rst genetic algorithm only. Of all the four methods for the object detection problems, the best one is the use of the back error propagation algorithm in the rst stage and the re ned genetic algorithm in the second stage.