Scholarlink

<p>Because of good feature map representation of deep learning, deep models are adapted in image segmentation tasks. Nevertheless, these models are limited to pixel-wise fitting of the map which leads to no accurate boundary determination. Among the earlier methods of image segmentation, the parametric active contour is one of the prominent methods that can accurately determine boundaries. However, it is sensitive to weak edges and initial contour. This article aims to apply the capabilities of parametric active contour in image segmentation, while overcoming its limitations by developing a new deep learning based vector field that performs well at low resolution. A synergy is formed between deep network and Active Contour Models (ACMs) to apply the strength of later and take advantage of the former to eliminate the challenges of ACMs by developing a new vector field named “Deep Vector Field (DVF)”. The modified Holistically-nested Edge Detection network, an edge detection-based deep model, and the kernel of Vector Field Convolution are incorporated to predict a DVF for parametric active contour. The model is applied to a couple of the relevant segmentation problems, namely fundus images’ optic disc and optic cup segmentations, separately. REFUGE, DRISHTI-GS, and DRION-DB public datasets are used for method’s evaluation. The results indicate that, despite the low input resolution, our method outperforms its counterparts for optic disc segmentation problem for all datasets, and performs well for optic cup segmentation problem using DRISHTI-GS dataset and relatively weak using REFUGE dataset in terms of Dice and Intersection over Union (IoU) evaluation metrics. Most existing methods rely on high-resolution inputs and perform poorly when such data is unavailable. This study addresses that limitation by enabling accurate segmentation at low resolutions, with reduced computational cost and faster processing. Overall, the newly developed DVF shows to be a promising vector field to overcome the limitations of many segmentation problems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Effective Deep Vector Field Design for Active Contour-Based Image Segmentation

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Maryam Hajrasooliha,
  • Ahmad Reza Naghsh-Nilchi

摘要

Because of good feature map representation of deep learning, deep models are adapted in image segmentation tasks. Nevertheless, these models are limited to pixel-wise fitting of the map which leads to no accurate boundary determination. Among the earlier methods of image segmentation, the parametric active contour is one of the prominent methods that can accurately determine boundaries. However, it is sensitive to weak edges and initial contour. This article aims to apply the capabilities of parametric active contour in image segmentation, while overcoming its limitations by developing a new deep learning based vector field that performs well at low resolution. A synergy is formed between deep network and Active Contour Models (ACMs) to apply the strength of later and take advantage of the former to eliminate the challenges of ACMs by developing a new vector field named “Deep Vector Field (DVF)”. The modified Holistically-nested Edge Detection network, an edge detection-based deep model, and the kernel of Vector Field Convolution are incorporated to predict a DVF for parametric active contour. The model is applied to a couple of the relevant segmentation problems, namely fundus images’ optic disc and optic cup segmentations, separately. REFUGE, DRISHTI-GS, and DRION-DB public datasets are used for method’s evaluation. The results indicate that, despite the low input resolution, our method outperforms its counterparts for optic disc segmentation problem for all datasets, and performs well for optic cup segmentation problem using DRISHTI-GS dataset and relatively weak using REFUGE dataset in terms of Dice and Intersection over Union (IoU) evaluation metrics. Most existing methods rely on high-resolution inputs and perform poorly when such data is unavailable. This study addresses that limitation by enabling accurate segmentation at low resolutions, with reduced computational cost and faster processing. Overall, the newly developed DVF shows to be a promising vector field to overcome the limitations of many segmentation problems.