| 摘 要: 针对遥感图像的语义分割存在边缘效果不佳以及遥感图像数据量庞大的问题,提出了一种结合超像素的轻量级语义分割模型LEDNet-SSMS。通过使用超像素 HQS(High Quality Superpixel Generation Through Regional Decomposition)对语义分割结果的边缘进行矫正,使得语义分割的结果具备HQS超像素强大的边缘一致性。使用超像素分割、合并、再分割的形式减少过分割和欠分割对整体修正效果的影响。LEDNet-SSMS在Massachusetts-building数据集上P值达到74.2%、IoU(Intersectionoverunion)达到70.4%、OA(Over Accuracy)达到91.2%;在FloodNet数据集上mPA(mean Pixel Accuracy)值达到87.2%、mIoU达到80.6%、OA达到94.9%,量化指标均高于其他算法,具有训练时间短、精度较高的优势。 |
| 关键词: 遥感图像 语义分割 区域合并 超像素 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 陕西省自然科学基金项目(2024JC-YBQN-0724) |
|
| LEDNet-SSMS:Lightweight Remote Sensing Semantic Segmentation Model Combined with Superpixel Model |
|
PAN Xinxin, YANG Xubing
|
(College of Information Science and Technology, Nanjing Forestry University, Nanjing 210037, China)
xinxin@njfu.edu.cn; xbyang@njfu.edu.cn
|
| Abstract: To address the issues of suboptimal edge performance in semantic segmentation of remote sensing images and the massive data volume of remote sensing imagery, this paper proposes LEDNet SSMS—a lightweight semantic segmentation algorithm incorporating superpixels. By utilizing HQS (High Quality Superpixel Generation Through Regional Decomposition) superpixels to refine the edges of semantic segmentation results, the algorithm ensures the segmentation outcomes inherit the robust edge consistency of HQS superpixels. The approach employs a sequence of superpixel segmentation, merging, and re-segmentation to mitigate the impact of ove-r segmentation and under segmentation on overall refinement. On the Massachusetts-building dataset, LEDNe-t SSMS achieves a Precision of 74.2% , IoU(Intersection over union) of 70.4%, and OA (Overall Accuracy) of 91.2% . On the FloodNet dataset, it attains mPA (mean Pixel Accuracy) of 87.2%, mIoU (mean Intersection over Union) of 80.6%, and OA of 94.9% . All quantitative metrics surpass those of other algorithms, demonstrating advantages in shorter training time and higher accuracy. |
| Keywords: remote sensing imagery semantic segmentation region merging superpixel |
