The Impact of Imbalanced Training Datasets on CNN Performance in Typical Remote Scenes Classification

Abstract

Convolution neural network (CNN) has a prominent performance in image classification. This thesis empirically studies how imbalanced training datasets impact on CNN performance in typical remote scenes classification. I adopted different strategy based on the complexity of scenes to set up imbalanced and balanced training datasets. Then I compared the results after training and testing experiments. Experimental results show that appropriate imbalanced strategy could improve the CNN performance in typical remote scenes classification. In face of specific classification tasks, the empirically method is a good choice to improve CNN performance as a whole.