A domain independent approach to multi-class object detection using genetic programming

Object detection is the process of automatically nding objects of interest within an image. Given the abundance of information now captured and stored in electronic form, this is fast becoming a useful and challenging machine learning and computer vision task. This project uses a domain independent genetic programming approach to solve four multi-class object detection problems ranging in dif culty levels. Three methods were investigated, all using different features sets based on local-region pixel statistics. A new measure, program size, was also introduced to the tness function with the aim of favoring the evolution of smaller programs over larger, complicated ones. The new tness function with program size proved more effective and ef cient, and the evolved programs using this tness function were much shorter and easier to interpret. While the classi cation method greatly reduced training times than the basic detection method, it could not improve the detection performance. The two-phase method with a secondary training phase always gave a better detection performance than without.