Evaluation of segmentation methods for RGB colour image-based detection of Fusarium infection in corn grains using support vector machine (SVM) and pre-trained convolution neural network (CNN)

Abstract

This study evaluated six segmentation methods (clustering, flood-fill, graph-cut, colour-thresholding, watershed, and Otsu’s-thresholding) for segmentation accuracy and classification accuracy in discriminating Fusarium infected corn grains using RGB colour images. The segmentation accuracy was calculated using Jaccard similarity index and Dice coefficient in comparison with the gold standard (manual segmentation method). Flood-fill and graph-cut methods showed the highest segmentation accuracy of 77% and 87% for Jaccard and Dice evaluation metrics, respectively. Pre-trained convolution neural network (CNN) and support vector machine (SVM) were used to evaluate the effect of segmentation methods on classification accuracy using segmented images and extracted features from the segmented images, respectively. The SVM based two-class model to discriminate healthy and Fusarium infected corn grains yielded the classification accuracy of 84%, 79%, 78%, 74%, 69% and 65% for graph-cut, watershed, clustering, flood-fill, colour-thresholding, and Otsu’s-thresholding, respectively. In pretrained CNN model, the classification accuracies were 93%, 88%, 87%, 84%, 61% and 59% for flood-fill, graph-cut, colour-thresholding, clustering, watershed, and Otsu’s-thresholding, respectively. Jaccard and Dice evaluation metrics showed the highest correlation with the pretrained CNN classification accuracies with R2 values of 0.9693 and 0.9727, respectively. The correlation with SVM classification accuracies were R2–0.505 for Jaccard and R2–0.5151 for Dice evaluation metrics.