Scholarlink

<p>The task of pixel-wise image segmentation remains significantly challenging in the domains of computer vision and image processing. This paper tries to address the complexities involved in spatial understanding for the semantic segmentation of high-resolution upright images by using Deep Convolutional Neural Networks. In this study, we worked with high-resolution images of substantial size. However, in order to facilitate smooth processing on a moderately configured computer with standard specifications, we cropped the images and reduced their size. We employed the Unet model for the network, which is frequently used for input formats that are identical to its intended function. With the continuous emergence of new bracket networks, it is crucial to recognize that the backbone performance of semantic segmentation networks may differ depending on the bracket network utilized. This study presents a comparative analysis of the performance variations among ResNet34, InceptionV3, and VGG16 when employed as the backbone for Unet, each exhibiting distinct strengths and weaknesses. Consequently, this opens the possibility of leveraging a combination of various base learners, which may surpass the performance of a singular segmentation model, given the pronounced sensitivity of deep learning models to differing network architectures. This paper outlines the creation of a weighted ensemble of multiple CNNs aimed at semantic segmentation of aerial images, harnessing the advantages provided by diverse CNN models to enhance both accuracy and generalization performance. The weighted ensemble algorithm was implemented by leveraging transfer learning on a Unet CNN model, with ResNet34, InceptionV3, and VGG16 serving as the encoder bases. Each U-Net model variant was trained using a unique pre-trained network, and their predictions were combined through a weighted average method aimed at multi-class semantic segmentation. The contribution of each model to the weighted average is dictated by its assigned weight. Our methodology outperformed individual deep learning models in independent evaluation, demonstrating improved classification accuracy. Additionally, a grid search algorithm was employed to optimize the assigned weights. Our experiments utilizing the MBRSC Dubai Aerial Imagery Dataset demonstrated both quantitative and qualitative enhancements, suggesting that our approach boosts segmentation accuracy and reduces distance errors compared to relying on a single classifier. The experimental results demonstrate that the weighted average ensemble achieved superior segmentation on aerial images, reaching 87.08% accuracy, 87.06% Dice, and 78.13% IoU, outperforming individual backbones, standard U-Net, FCN-8&#xa0;s, and advanced segmentation architectures such as, HRNet, and DeepLabV3 + .</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

Semantic segmentation performance of aerial image segmentation using weighted ensemble trained networks CNNs

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Zahra Faska,
  • Lahbib Khrissi,
  • Khalid Haddouch,
  • Nabil El Akkad

摘要

The task of pixel-wise image segmentation remains significantly challenging in the domains of computer vision and image processing. This paper tries to address the complexities involved in spatial understanding for the semantic segmentation of high-resolution upright images by using Deep Convolutional Neural Networks. In this study, we worked with high-resolution images of substantial size. However, in order to facilitate smooth processing on a moderately configured computer with standard specifications, we cropped the images and reduced their size. We employed the Unet model for the network, which is frequently used for input formats that are identical to its intended function. With the continuous emergence of new bracket networks, it is crucial to recognize that the backbone performance of semantic segmentation networks may differ depending on the bracket network utilized. This study presents a comparative analysis of the performance variations among ResNet34, InceptionV3, and VGG16 when employed as the backbone for Unet, each exhibiting distinct strengths and weaknesses. Consequently, this opens the possibility of leveraging a combination of various base learners, which may surpass the performance of a singular segmentation model, given the pronounced sensitivity of deep learning models to differing network architectures. This paper outlines the creation of a weighted ensemble of multiple CNNs aimed at semantic segmentation of aerial images, harnessing the advantages provided by diverse CNN models to enhance both accuracy and generalization performance. The weighted ensemble algorithm was implemented by leveraging transfer learning on a Unet CNN model, with ResNet34, InceptionV3, and VGG16 serving as the encoder bases. Each U-Net model variant was trained using a unique pre-trained network, and their predictions were combined through a weighted average method aimed at multi-class semantic segmentation. The contribution of each model to the weighted average is dictated by its assigned weight. Our methodology outperformed individual deep learning models in independent evaluation, demonstrating improved classification accuracy. Additionally, a grid search algorithm was employed to optimize the assigned weights. Our experiments utilizing the MBRSC Dubai Aerial Imagery Dataset demonstrated both quantitative and qualitative enhancements, suggesting that our approach boosts segmentation accuracy and reduces distance errors compared to relying on a single classifier. The experimental results demonstrate that the weighted average ensemble achieved superior segmentation on aerial images, reaching 87.08% accuracy, 87.06% Dice, and 78.13% IoU, outperforming individual backbones, standard U-Net, FCN-8 s, and advanced segmentation architectures such as, HRNet, and DeepLabV3 + .