Small object detection in highly variable backgrounds

Abstract

The analysis of imagery from outdoor remote sensing is a technique widely used for surveying and data gathering. This paper studies techniques to be deployed in small object localisation using Convolutional Neural Networks (CNN), with the aim to detect litter in outdoor non-urban imagery. The detection of small objects requires distinguishing features between foreground and background. A litter detection application has to counter high variability in the foreground, as litter is defined as a super-class of common objects, and the high variability found in a rural or coastal backgrounds. Remote sensing imagery of non-urban scenery does not offer high contrasting features, reducing the effect of normal object localisation techniques.