LEARNT HORIZON FILTER: A MACHINE LEARNING APPROACH

Abstract

: In digital image, the sky colour detection has number of applications. It includes, but not limited to lighting correction, image enhancement, horizon alignment, scene indexing and others. This article tackles the problem of pixel based sky colour detection from the machine learning point of view. Rather than creating complex filters, the setup in this article uses simple pixel classification approach by the offline trained classifiers. From the machine learning set, four classifiers are used, including: Random Forests, Multi-layer Perceptron, Radial Basis Function and the Bayesian Network. The experimental evaluation is presented on a dataset of 1000 images. Experimental results show the feasibility of the Multilayer Perceptron for sky detection. It is also found that the Multilayer Perceptron classifier has 8% higher detection performance compared to the Random Forest classifier and the Radial Basis function classifier. The Random Forest classifier however has 9% higher performance compared to the Bayesian classifier and approximately equal to the Radial Basis function. This research work not only presents colour based sky detection, but also contributes and benefits the colour based object detection in general