Automated Detection of Root Crowns using Gaussian Mixture Model and Bayes Classification

In this paper a method for automatic detection of root crowns in root images, are designed, implemented and quantitatively compared. The approach is based on the theory of statistical learning. The root images are preprocessed with algorithms for intensity normalization, segmentation, edge detection and scale space corner detection. The features used in the experiments are the Zernike moments of the bi-level image patch centered around high curvature detections. Zernike moments are orthogonal and thus can be rightly assumed to be independent. The densities of the feature vectors for different classes are modelled with Gaussian mixture model (GMM), with a diagonal covariance matrix. The parameters for the feature's distribution densities for different classes are learnt by expectation maximization. Bayes rule and Neymann-Pearson criteria is used to design the classification method. We experiment with different orders of Zernike moments and different number of Gaussians in the GMM. The experiments are done on a real dataset with images of rice, corn, and grass roots. Pattern classification results are quantitatively analyzed using Receiver Operating Characteristic (ROC) curves and area under the ROC curves. We quantitatively compare the results of the proposed method with that of support vector machine (SVM) which is another very popular statistical learning method for pattern classification.