IMAGE CLASSIFICATION WITH DEEP AND ENSEMBLE LEARNING APPROACHES
Project no. 1328 · Status: Kapanmış · 2018
Summary
Deep Learning is a sub-field of Machine Learning, and when more examples are used, higher performance can generally be achieved compared to traditional methods. The Ensemble Learning approach, on the other hand, aims to increase performance by combining two or more machine learning models in solving a problem. In our project, 3 new approaches have been proposed regarding the use of Deep Learning and Ensemble Learning approaches separately or together in the field of Image Classification. In the first study, the features extracted from the AlexNet deep learning model were combined with the manually obtained ones, and the accuracy value on the CIFAR-10 dataset could be increased with the voting-based ensemble classification approach. In the second study, the performance values in CIFAR-10 and CIFAR-100 image datasets were increased by combining the features obtained from pre-trained transfer learning models and other fine-tuned convolutional neural network models. As another contribution, adding lower-dimensional layer to the end of some models has also been observed to have a positive effect on the result. In our last study, we propose voting, data augmentation, and preprocessing variations-based novel method employing InceptionV3 and Xception transfer learning models for Tuberculosis detection. After 40 model combinations are generated, the best ones are selected and used for voting operation as base classifiers in the ensemble learning phase. The Accuracy and Area Under the ROC Curve (AUC) values were reached on the Montgomery County and Shenzhen TB CXR datasets, which compete with the best results in the literature.