Author: Mengjie Zhang
Source: GZipped PostScript (997kb); Adobe PDF
(1002kb)
We describe a pixel based approach to the use of neural networks for multiclass object detection problems in which the classes and locations of relatively small objects in large pictures must be determined. The networks use a squared input field which is large enough to contain every single object of interest and are trained by the back propagation algorithm on examples which have been cut out from the large pictures. The trained networks are then applied, in a moving window fashion, over the large pictures to detect the objects of interest. This approach has been examined on three object detection problems of increasing difficulty. The results suggest that this approach can be used to detect simple and regular objects with the translation and rotation invariance in large pictures against a relatively uniform background.