Weed Pixel Level Classification Based on Evolving Feature Selection on Local Binary Pattern with Shallow Network Classifier

Abstract

This paper proposes an evolving feature selection on a rotation-invariant Local Binary Pattern (LBP) with genetic algorithm (GA) and a non-linear classifier to perform pixel-wise classification on biomass pixels. Early true leaf growth of carrots and weeds consists of 60 images provided by a public benchmark Crop/Weed Field Image Dataset (CWFID) [1] was used. The GA encodes LBP feature parameters generated from normalized difference vegetation index (NDVI) image slices as genome consisting of number of neighbours, radius, size of image slice, number of LBP combinations. LBP is a lightweight rotation invariant texture feature descriptor which encodes discriminative local texture information between crops and weeds. The study evaluated multiple ensemble models evolved by GA, where GA evolves the LBP feature parameter selection, and the number of LBP features as input variables used. The classifier can handle crop and weed overlaps, which are often present in commercial fields. Weeds often grow in close proximity and overlap crops and are similar in size, which adds complexity in discriminating them. Our experiments have shown that combinations with a back-propagation neural network with a symmetrical two hidden-layer network achieved the best classification accuracy when compared to other non-linear classifiers. By utilizing a single type of feature (LBP texture feature), our resulting Artificial Neural Network (ANN) ensemble classifier is able to achieve a classification accuracy of 83.5 %.