Insect Event Extraction in LIDAR Images using Image Analysis and Convolutional Neural Networks
Kontakt: magnuso [at] maths [dot] lth [dot] se
Spara händelsen till din kalender
Ludvig Malmros presenterar sitt examensarbete "Insect Event Extraction in LIDAR Images using Image Analysis and Convolutional Neural Networks"
Abstract:Insect monitoring has earlier been a manual, tedious and time consuming task that is impossible to do in real time. Thus there exists a need for a real time automatic insect monitoring system for counting and classifying insect for use in scientiﬁc research and pesticide spraying control. One approach to automatize this is using LIDAR to detect insects. In this thesis it has been explored how to detect, segment and do merge classiﬁcation on insect events showing up in large 2D time-range map frames created from a LIDAR optics setup as insects ﬂy through a laser beam. The suggested extraction method combines simple intensity thresholding techniques and well known edge detection techniques, together with a statistical, background-noise based, dilation iteration stopping criteria for region growing. This is used to segment insect events while avoiding accidental splitting and merging of events. Thereafter a trained convolutional neural network is suggested to classify all events that might have been merged, such that they could be discarded instead of being inputted to the species classiﬁcation system. Tests and observations indicate that the old segmentation method ﬁnds close to all the wanted insect events, but over-segments them drastically in some cases. By dividing the method into one event detection and one border ﬁnding part, the suggested extraction method are able to ﬁnd the same amount of events without increasing the number of splits and merges. At the same time it is able to ﬁnd event segmentation borders with a higher precision then previously possible. Tests on the merge classiﬁcation indicate surprisingly good results for the ability to classify event merges. Creating artiﬁcial merges to handle the imbalanced data set shows further improvement, while oversampling does not. Indicated is also that the size compression used does not seem to eﬀect the classiﬁcation negatively.