Aerial Image Classification Using Convolution Neural Network

Abstract

Image classification task has been an important area in the field of computer vision study as it is applied in varied applications. The approaches for image classification is based on feature selection of image class and effectively applied through various low level feature algorithms for matching with a particular class and yielding a classification result identifying one from another. The task for feature selection particularly become extremely challenging to spatial images such as aerial images, as they contain varied complex feature, scale challenges as well as image orientation. A more accepted approach of image classification for aerial image is the use of Convolution Neural Networks (CNN). CNN models are capable of producing higher accuracy compared to contemporary feature based algorithm as they tend to utilize higher local features. Transfer learning approach using readily available and proven CNN for image classification makes the task a step closer to capturing and designing a CNN which adapt to user dataset and classification requirements. There is however the need to create user dataset with sufficient images to retrain the network for fine-tuning and testing its accuracy. This paper presents such experiment using GoogLeNet pre-trained network which is subject to replacing of the fully-Connected layer and output layer as per the classification requirement. The network is further subject to training and testing using test dataset other than that have never been exposed to the network. An accuracy of 98.33% was achieved.