The core purpose of this thesis is the comparison between the three most used novelty detection methods in the literature namely, Gaussian mixture models, one-class support vectors machine and k-nearest neighbors. Also different feature extraction methods are discussed and compared. The three methods are trained on the GAPs data set (The German Asphalt Pavement Distress), Eisenbach et al., 2017, using only the intact data with the goal of detecting distress in the images. GAPs is the first freely available pavement distress dataset providing high quality images recorded by a standardized process fulfilling German federal regulations. This will provide a base for a fair comparison of researches in this field.
- Gabor Filter
- Gray-level co-occurrence matrix
- Local Binary Pattern
The structure is divided as follows:
- Produce-all-results: where the actual training and evaluation of the three models occur (.py files)
- Notebooks-FeatureExtraction : The notebooks for generating features from training, validation and test datasets.
- Extracted Features: numpy arrays of all extracted features using the three aforementioned methods.
- All Notebooks: contain all experimental notebooks that was written during the thesis (Here you can execute the code to see ROC curves).
Performance of Gaussian mixture model using the three different feature extraction methods on Validation set
Performance of GMM, SVM, KNN using LBP as the feature extraction method on Validation set
