Integrating Contour Detection And Grab-Cut For Improved Harumanis Leaf Image Segmentation

Nurul Akmar Azman, Fatimah Khalid, Puteri Suhaiza Sulaiman, Zainal Abdul Kahar
dx.doi.org/10.17576/apjitm-2025-1402-17

Abstract

Accurate disease detection on Harumanis mango leaves is often hindered by variations in lighting conditions and inconsistent visual features such as differences in color intensity, textures, and the presence of shadows, that may complicate the segmentation process. This study introduces the Contour-Gamma-Grab Cut (CGGC) method, a novel image processing technique designed to enhance background removal to improve segmentation?s accuracy. The CGGC method integrates contour detection, gamma adjustment, and the Grab-Cut algorithm to tackle these challenges. We systematically compare CGGC against classical thresholding methods such as Sauvola and Niblack, demonstrating its superior performance. Additionally, we assess the efficacy of CGGC in combination with deep learning architectures such as 3 layers Convolutional Neural Network (CNN), U-Net, and SegNet. Results show that CGGC, when combined with handpicked segmentation algorithm such as normal thresholding, achieves a specificity of 97.04% and a sensitivity of 96.65%, outperforming existing methods. Moreover, CGGC-processed images analyzed with U-Net achieve specificity and sensitivity scores of 99.99% and 98.99%, respectively when trained with 100 epochs, marking unprecedented accuracy in segmentation. These findings underscore CGGC's potential to enhance disease detection accuracy in agricultural contexts, particularly for Harumanis mango leaves.

keyword

Grab-Cut, Contour Adjustments, Segmentation, Deep Learning, Thresholding.

Area

Pattern Recognition