Pixel Statistics Based Neural Networks for Domain Independent Multiclass Object Detection


Authors: Mengjie Zhang, Peter Andreae, Roy Chow
Source: GZipped PostScript (919kb); Adobe PDF (396kb)

The paper describes a domain independent approach to multiclass object detection problems in which the classes and locations of relatively small objects in large images must be determined. The approach uses low-level pixel statistics rather than domain dependent features as inputs to neural networks. The networks are trained by the back propagation algorithm on examples which have been cut out from the large images. The trained networks are then applied, in a moving window fashion, over the large images to detect the objects of interest. The paper reports on a series of experiments on three object detection problems of increasing difficulty, and with different sets of pixel statistics. The results suggest that this approach can be used successfully to detect simple and regular objects in large pictures against a relatively uniform backgrounds. Using moment statistics are local region pixel statistics improves multiclass object detection performance significantly.

[Up to Computer Science Technical Report Archive: Home Page]