Comparative Evaluation of Segmentation Techniques on Maize Seeds

Abstract

In this study, we performed a comparative analysis of various segmentation techniques on a dataset of digital images depicting maize seeds. The dataset includes categories representing different quality levels of seed batches, labeled as worst, bad, average, good, and excellent. The objective of this study was to identify the most accurate combination of segmentation techniques and color spaces for maize seed segmentation. To achieve this, we systematically compared various segmentation techniques, including Otsu, Watershed, and K-means clustering, across different color spaces such as RGB, HSV, and grayscale.

We evaluated the performance of these combinations using metrics like Intersection over Union (IoU), Mean Square Error (MSE), Pixel Accuracy (PA), Kappa (k) coefficient, and F-1 Score in images from the Maize Seed dataset. The results showed that the K-means segmentation technique in the HSV color space yielded the best results across all evaluated metrics. In particular, using the IoU metric in the HSV color space, K-means achieved mean values of 0.98, 0.97, 0.95, 0.91, and 0.91 for the categories excellent, good, average, bad, and worst, respectively.