IMPROVEMENT OF DEEP LEARNING MODELS ON CLASSIFICATION TASKS USING HAAR TRANSFORM AND MODEL ENSEMBLE

Abstract

Machine learning have an enormous impact on Computer Vision. This thesis investigates how to improve efficiency of a Machine learning technique called deep learning on classification tasks using Haar transform and model ensemble. Haar transform can be used to reduce the input size of image data to train more models and model ensemble is known to boost performance using multiple models instead of one.

Experimental results showed that Adaboost and stacking work as expected as they boosted the accuracies by 2-3% and approximately 1%. However, the other two methods, averaging and geometric mean, did not boost but scored between the best individual model and the worst. This thesis also suggests future work onto Adaboost and stacking.