Convolutional Neural Network (CNN) features behaviour in the context of textures
|Title||Convolutional Neural Network (CNN) features behaviour in the context of textures|
|Summary||The project aims to quantify the behaviour of Convolutional Neural Network (CNN) features in the context of textures.|
|Prerequisites||Image Analysis: Grade 4 or 5|
|Supervisor||Josef Bigun, Kevin Hernandez-Diaz, Fernando Alonso-Fernandez.|
First it will Investigate how the cnn's behave when the number of classes increase, potentially to infinity. Here, syntethic images can be used since the groundtruth is known. One such classificaiton problem comprises finding the local orientations on a series of Frequency Modulated test images and associated certainties. The number of classes can be made as large as one wishes in both cases, the discretizaton grid being the limit (In FM-test type images this equals to their sizes).
Second, Convolutional Neural Network (CNN) features discrimination power will be be quantized in terms of within and between classsvariances, with different textures in a mosaic of texture images, cut from real aerial images and Brodatz textures.