Navigation auf uzh.ch
Navigation auf uzh.ch
Before the deep learning boom, most image processing algorithms relied on gray scale images since the handling of color was too complex and often not required to fulfill the task at hand. Examples for such image processing algorithms are the Scale Invariant Feature Transform (SIFT), Local Binary Patterns (LBP) or Gabor features.
With the introduction of deep learning, people immediately started using color inputs as there is a strong believe that color helps. For some tasks like classifying a ripe from an unripe fruit, color is surely a useful feature, but it is not clear that color helps in general and for all tasks.
Another question is how color is handled in deep networks. Currently, the fusion of the three color planes (R,G,B) is done already in the very first layer of the network and researchers never questioned this behavior. However, due to the nonlinear nature of the network it is unclear if it would maybe better to learn convolutional filters on the color channels separately and fuse those channels later in the network structure.
The task in this Bachelor's thesis is to investigate how much the current usage of color information in deep learning influences classification accuracy. Different network topologies must be engineered and tested against well-known image classification tasks such as CIFAR-10 and if time allows against the larger ImageNet Dataset \cite{russakovsky2015imagenet}.
This topic can also be turned into a Master's thesis. In that case, additional experiments will need to be performed in order to test Face Recognition networks.
