Object recognition using geometric properties
This thesis presents an improved method for object recognition in grayvalue images. An earlier method used the boosting algorithm AdaBoost for learning a simple recognizer. The recognizer labels an image as positive, if certain simple features appear in the image. Geometric relations between simple features were ignored. In this work the learning algorithm LPBoost is used to improve recognition results. For further improvement simple geometric relations between features are used. The search for geometric relations is implemented within the weak learner, which produces the weak hypotheses for the boosting algorithm. A full search for relevant geometric relations between simple features is rather impossible because of computation time. This work therefore propose two greedy search strategies. The system is tested with difficult real-world datasets. The experiments show that using geometric relations gains an advantage over the simple method, if thedataset is difficult. A comparison between the two boosting algorithms is also performed. It shows that the LPBoost algorithm is a clear winner over AdaBoost.