A Network Model for Fragment-Based Object Classification Master Thesis

The fragment-based approach for visual object recognition and classification suggests that informative image fragments are useful for performing classification and recognition tasks. Informative fragments proved to be efficient features for object classification in computational studies of this approach. Biological studies support the possibility that visual recognition processes use such fragments. This thesis supports the possibility that informative fragments can be learned and used for these tasks by a biological system. This is shown by presenting a neural network model for fragment-based classification. The proposed neural network model extracts informative image fragments from image examples and uses them for classification. A key feature of the model is that it uses " neuronal imprinting " , in which a receptive field of a neuron can be determined by a single presentation. Using this in the training process, single neurons develop sensitivity to image fragments, and the informative ones are selected through competition between the neurons. The work experimentally shows that the performance of the proposed model in classification tasks is similar to the performance of computational methods that use informative image fragments. Finally, the biological plausibility of the network model is discussed, and the biological mechanisms required for using informative fragments in a biological system are identified.