Identification of small objects using convolutional neural networks : a case study of litter detection from aerial imagery

Abstract

The use of remote sensing is taxing on Convolutional Neural Network (CNN) object localisation performance because out door imagery provides non-ideal image scene capture that reduces the image quality of the object to be detected. The task is rendered more arduous when the objects to be detected are found in rural and coast line terrains, where the background image(orterrain) is variable and does not provide high foreground/background discriminatory features. This is a case study using CNN algorithms to detect litter from high resolution imagery obtained from drone surveys from such terrain. The study describes a method how to use off-the-shelf inference engines that are trained in a relatively small amount of time, using pre-trained weight sand fine tuned on designed datasets, whilst studying performance effects of data augmentation, normalisation and image pre-processing techniques. The sensitivity to high variability in the terrain was reduced by implementing training-dataset engineering. The use of Class Activation Mapping and Overlap suppression performs a weak localisation technique providing more defined litter localisation, improving Intersection over Union (IoU) values. A drone survey dataset was compiled from images obtained from 12 land surveys and litter annotation sessions. A small object algorithm average precision of 0.38 achieved at IoU:0.1 and when taking into account the small dimensions of the objects obtaining an AP of 0.53 at IoU:0.01.