Railway wagons classification

This Master's thesis deals with classification of railway wagons based on visual information. A theoretical background of two different approaches for a classification system is provided and both approaches are subsequently implemented. First approach includes transforming images of wagons to histograms of visual words, according to the Bag of Visual Words method. Afterwards, classifiers such as k-NN, SVM, Multinomial Naive Bayes, neural network and Ensemble method, specifically Voting classifiers, are applied. Second approach is classifying images using well known architectures of Convolutional Neural Networks and transfer learning. AlexNet, VGG16 and ResNet50 were pre-trained on a large ImageNet dataset and the upper layers were trained on the dataset of railway wagons. Both approaches were fine-tuned for the best possible performance. For comparison of both approaches a training dataset with 1773 images in 27 classes and testing dataset with 444 images were compiled. On testing dataset the best classifier using BoVW method reached accuracy of 89%. Convolutional neural nets performed with 95-97% accuracy, which is an improvement. Prediction times of images to be classified are also considered. Beyond the scope of the assignment of this thesis, an algorithm for splitting train images into images of individual wagons was developed. In the conclusion, limitations and reasons for limited robustness of this algorithm are presented