Search

The research work conducted in the frame of this doctoral thesis does satisfactorily solve a series of research questions, which do refer to four different and specific technical contexts, i.e. Context-1 “house-types classification”, Context-2 “document-image classification”, Context-3 “document-image enhancement”, and Context-4 “Robust Optical Character Recognition under difficult conditions”, which integrates both text detection and text classification.Consider Context-1 (related to solving selected computer vision tasks using machine learning w.r.t. a robust “house-types classification”). The following innovative contributions have been developed and comprehensively validated. Hereby, a novel visual sensing concept for robustly classifying house types through a convolutional neural network architecture involving a multi-channel features extraction has been developed, comprehensively validated, and benchmarked with selected best-competing approaches from the relevant literature. The superiority of the novel concept has been well demonstrated.Consider Context-2 (related to solving selected computer vision tasks using machine learning w.r.t. a robust “document-image classification”). The following innovative contributions have been developed and comprehensively validated. Indeed, a novel hybrid convolutional neural network model for a robust classification of document-Images under real-world hard conditions has been developed, comprehensively validated, and benchmarked with selected best-competing approaches from the relevant literature. The superiority of the novel concept has been well demonstrated. Consider Context-3 (related to solving selected computer vision tasks using machine learning w.r.t. a robust “document-image enhancement”). The following innovative contributions have been developed and comprehensively validated. Indeed, a novel patch-based convolutional neural network model for blind blur enhancement of document-images has been developed, comprehensively validated, and benchmarked with selected best-competing approaches from the relevant literature. The superiority of the novel concept has been well demonstrated. Consider Context-4 (related to solving selected computer vision tasks using machine learning w.r.t. a robust “Optical Character Recognition under difficult conditions”). The following innovative contributions have been developed and comprehensively validated. Indeed, two novel hybrid convolutional neural networks, a recurrent neural network model involving Residual Block layers, Gabor preprocessing layers for detecting of the text boundary boxes, and a text recognition deep-neural model have been developed, comprehensively validated, and benchmarked with selected best-competing approaches from the relevant literature. The superiority of the novel concept has been well demonstrated., by Kabeh Mohsenzadegan, Dissertation Universität Klagenfurt 2022