Deep Learning has recently become a very popular topic in machine learning due to its successes in several areas. These include image processing, natual speech recognition, artificial intelligence in computer games, and many more! Even the big companies as Microsoft, Google, Facebook, Apple put lots of effort in improving the methods of Deep Learning to create machines that reach human performances on specific tasks.
Deep Learning is based on neuronal nets that consist of many layers and their connections. The fundamentals for training these networks e.g Backpropagation are already known since 1986 by the work of D. Rumelhart, G. Hinton, and R. Williams, but they could not be applied to deep networks because of the lack of computing power. This is the main reason why training those networks is usually done on GPUs.
An other important requirement for Deep Learning is the presence of large datasets that contain many different examples for the same class to prevent overfitting. Usually one uses data augmentation to extend a dataset by creating new examples out of the existing ones. Therefore one applies random label preserving operations on the images, e.g. scaling, rotations, changes in contrast, color.
We apply the Deep Learning techniques especially on image processing tasks using a special Convolutional Neuronal Network (CNN) structure.
We use CNNs to classify images on the famous Flavia, Foliage and MEW dataset, which contain images of leaves of different tree species. The goal is to predict the species of a leaf.
This is a challanging task because there exist many leaves that differ only marginally, which is why it is tough to see differences even for humans.
By applying data augmentations on the rather small datasets we achieve accuracies of over 99% on this task, which is competitive in comparison with methods that are based on feature extractions.
Segmentation of medical data
A primary goal of artifictial intelligence in medical applications is to assist a doctor in medical outcomes, e.g. by detecting anomalies in radiographs. As first step a segmentation of the different organs of an image is usually required. An example outcome shows the image on the right hand side: the lung is segmented in blue and the heart in red. This image was produced by usage of a CNN-based pixel classifier that labels each pixel of the image depending on a window around this pixel.
Another ongoing task is to detect cancer in liver tissue. For accessing the required medical data we work with the University Clinic Würzburg.
Live dataset creation and augmentation tool
The standard technique for aumenting a dataset is to create additional examples before training in a fixed way, i.e. one extends the number if examples by a fixed amount. We develop a tool that allows to generate images live during the training which allows to generate a so to say infinite augmentation of the initial dataset.
Moreover this tool allows to automatically train and predict pixel classifier tasks and postprocess the produced image.
If you are interested in Deep Learning or want to get involved with our group contact C. Wick. We offer Bachelor and Master Thesis or master practica in this topic.