Enhancement of the Fynbos Leaf Optical Recognition Application (FLORA-E)

Object perception, classification and similarity discernment are relatively effortless tasks in humans. The exact method by which the brain achieves these is not yet fully understood. Identification, classification and similarity inference are currently nontrivial tasks for machine learning enabled platforms, even more so for ones operating in real time applications. This dissertation conducted research on the use of machine learning algorithms in object identification and classification by designing and developing an artificially intelligent Fynbos Leaf Optical Recognition Application (FLORA) platform. Previous versions of FLORA (versions A through D) were designed to recognise Proteaceae fynbos leaves by extracting six digital morphological features, then using the k-nearest neighbour (k-NN) algorithm for classification, yielding an 86.6% accuracy. The methods utilised in FLORA-A to -D are ineffective when attempting to classify irregular structured objects with high variability, such as stems and leafy stems. A redesign of the classification algorithms in the latest version, FLORA-E, was therefore necessary to cater for irregular fynbos stems. Numerous algorithms and techniques are available that can be used to achieve this objective. Keypoint matching, moments analysis and image hashing are the three techniques which were investigated in this thesis for suitability in achieving fynbos stem and leaf classification. These techniques form active areas of research within the field of image processing and were chosen because of their affine transformation invariance and low computational complexity, making them suitable for real time classification applications. The resulting classification solution, designed from experimentation on the three techniques under investigation, is a keypoint matching – Hu moment hybrid algorithm who`s output is a similarity index (SI) score that is used to return a ranked list of potential matches. The algorithm showed a relatively high degree of match accuracy when run on both regular (leaves) and irregular (stems) objects. The algorithm successfully achieved a top 5 match rate of 76% for stems, 86% for leaves and 81% overall when tested using a database of 24 fynbos species (predominantly from the Proteaceae family), where each species had approximately 50 sample images. Experimental results show that Hu moment and keypoint classifiers are ideal for real time applications because of their fast-matching capabilities. This allowed the resulting hybrid algorithm to achieve a nominal computation time of ~0.78s per sample on the test apparatus setup for this thesis. The scientific objective of this thesis was to build an artificially intelligent platform capable of correctly classifying fynbos flora by conducting research on object identification and classification algorithms. However, the core driving factor is rooted in the need to promote conservation in the Cape Floristic Region (CFR). The FLORA project is an example of how science and technology can be used as effective tools in aiding conservation and environmental awareness efforts. The FLORA platform can also be a useful tool for professional botanists, conservationists and fynbos enthusiasts by giving them access to an indexed and readily available digital catalogue of fynbos species across the CFR.

Grantor

Department of Electrical Engineering

Year dc:date.issued

2021

Author dc:creator

• Makumborenga, Roy

Advisor dc:contributor.advisor

• Winberg, Simon

• Electrical Engineering

Handle dc:identifier.uri

http://hdl.handle.net/11427/33853

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/33853