Multi - class Object Classification and Detection Using Neural Networks

Two problems in computer vision are object classification and detection. Object classification is the determination of what category an object belongs to and object detection is the determination of where suspicious objects are in a large picture and what class they belong to. Given the advantageous of an automated recognition system, a solution to this problem has always been a desirable objective. This project investigates the application of two domain independent approaches to solve four multi-class object recognition problems ranging in difficulty. Using a pixel statistics and a raw pixel values based approach with a neural network within a recognition system, a preliminary methodology was applied. Results were promising, showing that using concentric local region pixel statistics for object classification problems, can outperform a raw pixel values based approach. A powerful new algorithm, the Donut algorithm, is introduced as a false alarm filter to improve the detection performance and results illustrate a markedly significant improvement when used with pixel statistics. The use of a more extensive training set is also investigated and results indicate that these can also significantly improve the performance when used with a raw pixel values based approach.