Flower image segmentation based on color analysis and a supervised evaluation

We propose a flower segmentation schema which overcomes some limits of previous works. Indeed, it does not involve any interaction with the user, or make assumptions based on the domain knowledge. To achieve segmentation, we used OTSU thresholding on Lab color space. The thresholding was performed, separately, on the three component L, a and b, and the best result is selected relatively to the ground truth. The experimentation of the proposed method, performed using the dataset from the Oxford flower collection, make better the results, while consuming less CPU time, than the method proposed by Nilsback and Zisserman[5].