Effect of Blurring on Identification of Aerial Images Using Convolution Neural Networks

Abstract

The emergence of deep learning in the field of computer vision has led to extensive deployment of convolutional neural networks (CNNs) in visual recognition systems for feature extraction. CNNs provide learning through hierarchical inferencing by providing multilayer architecture. Due to high processing capability of CNNs in multidimensional signals like images, they are considered to be predominant artificial neural networks. CNNs are extensively used in computer vision such as in image recognition where the intent is to automatically learn features followed by generalization and eventually recognizing the learned features. In this paper, we investigate the efficiency of CNNs: AlexNet and GoogLeNet under the effect of blurring which occurs frequently during image capturing process. Here, Gaussian blurring is used since it minimizes the noise embedded into the image. For experimentation, UC Merced Land Use aerial dataset is used to evaluate CNNs’ performance. The focus is to train these CNNs and classifying an extensive range of classes accurately under the influence of Gaussian blurring. Accuracy and loss are the parameters of classification considered for evaluating the performance of CNNs. Experimental results validated the susceptibility of CNNs towards blurring effect with GoogLeNet being more fluctuating to varied degrees of Gaussian blurring than AlexNet.