Pixel Statistics and Program Size in Genetic Programming for Object Detection


Authors: Mengjie Zhang, Urvesh Bhowan
Source: GZipped PostScript (644kb); Adobe PDF (284kb)

This paper describes an approach to the use of genetic programming for object detection problems. In this approach, domain independent, local region pixel statistics are used to form three terminal sets. The function set is constructed by the four standard arithmetic operators and a conditional operator. A multi-objective fitness function is constructed based on detection rate, false alarm rate, false alram position and program size. This approach is applied to three object detection problems of increasing difficulty. The results suggest that the concentric circular pixel statistics are more effective than the square features for these object detection problems. The fitness function with program size is more effective and more efficient for these object detection problems and the evolved genetic programs using this fitness function are much shorter and easier to interpret. Keywords: Genetic programming, pixel statistics, false alarm position, program size, multiclass object detection.

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