|Supervisor:||M.Sc. Martin Pöllot (Room 06.020)|
|Faculty:||Prof. Dr.-Ing. André Kaup|
Convolutional Neural Networks (CNNs) represent the best tool for classification of image content. Over the past few years, this area of research brought great progress to image classification. One of the most significant breakthroughs in the beginning is the reliant classification of handwritten postal zip numbers and later the recognition of faces or license plates. Current state of the art applications use powerful real-time-capable networks that are able to detect multiple classes in images for detecting pedestrians, vehicles, obstacles and traffic signs in real-time. Currently, methods for preprocessing of to be analyzed images are sparse. Rotating and flipping of the input data or cropping the input image to a random square partition of the input image are common preprocessing techniques. Additionally, all pixel values are shifted to normalized intervals. A plethora of other preprocessing methods are possible but not researched. A CNN is rated by its overall ability to classify. Ideally, preprocessing of input data increases the ability of the network to classify given input data more reliably.
Ms Heimann is presented with the task of examining the consequences of suitable preprocessing techniques for the input data on the classification results of a given CNN. The selection of preprocessing methods are the result of an in detail study of the state of the art. The implementation should be done on multiple provided architectures of the network. The main part of the thesis consists of implementing required adaptions to the CNN for differently developed preprocessing methods to eliminate systematic errors. Afterwards, the input data shall be processed accordingly and the effects of the preprocessing should be analyzed. Finally, well-founded decisions on the configuration of preprocessing techniques should be given.
The current state of the art shall be determined by conducting a literature research. The thesis includes a well-documented presentation of the results; any source code created shall include sufficient annotation.