Vision-Based Flowering Plant Detection and Recognition

Abstract

The problem of flowering plants recognition is considered as a challenging problem and it has been partially solved under a certain condition that the flowers be pre-detected and nearly optimal segmented. In this thesis, we study the flower detection and recognition approaches and we demonstrate the ability to detect flowers in full scene images by means of machine learning. State of the art feature extraction methods are considered for the detection problems including: Discrete Wavelet Transform, Histogram of Oriented Gradients and Gabor Filter. We consider the flowering plant detection and recognition as classification problems. A set of Linear Support Vector Machines is used for classification. Poslets as a detection approach is considered for flowering plants detection. For training the classifiers to recognize a flower, we use a set of benchmark images, the classifiers run over a natural test images using a multi-scale scanning window to find the strong activations. For full scene detection, We have built a new testing dataset of flowering plant species that include 10 flowering plant species. The experiments show that Discrete Wavelet Transform features as input for linear Support Vector Machine is superior in the performance of other features. We have achieved an accuracy rate in detecting 10 flower species reaches about 88%. Experimental results using Receiver Operating Characteristics viii