| 摘 要: 针对DeepLabV3+模型在土地覆盖分类遥感地物分割中存在的边界模糊、小目标漏检及样本分布失衡导致精度下降等问题,提出一种改进的轻量级神经网络FAD-Net。首先,将轻量级网络MobileNetV2替代原始Xception作为主干网络以降低参数量;其次,提出多分支深度可分离空洞空间金字塔(MBDC-ASPP)架构,在保持多尺度上下文捕获能力的同时降低模型复杂度;最后,通过双路注意力机制与特征融合模块提升跨尺度特征分割能力。将FAD-Net模型在LoveDA数据集上进行测试。实验数据表明:实验结果表明,相较于原始DeepLabV3+模型,本文提出的FAD-Net在LoveDA数据集上的平均交并比(mIoU)提升了4.34个百分点。通过轻量化网络结构设计,参数量相较原始DeepLabV3+模型仅有4.162×106,验证了算法在保持语义分割精度的同时,显著提升了计算效率。 |
| 关键词: 土地覆盖分类 轻量化 DeepLabV3+ 语义分割 |
|
中图分类号: TP391
文献标识码: A
|
|
| Research on Land Cover Classification Method Based on Improved Lightweight Semantic Segmentation Network |
|
LIN Anyuan, WU Lili
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
15550480135@ 163.com; wull@ gsau.edu.cn
|
| Abstract: To address the issues of blurred boundaries, missed detection of small targets, and accuracy decline due to imbalanced sample distribution in the land cover classification of remote sensing objects using the DeepLabV3+ model, this paper proposes an improved lightweight neural network, FAD-Net. Firstly, the lightweight network MobileNetV2 is used instead of the original Xception as the backbone network to reduce the number of parameters. Secondly, a mult-i branch deep separable dilated spatial pyramid ( MBDC-ASPP) architecture is proposed, which maintains the ability to capture mult-i scale context while reducing the model complexity. Finally, the cross-scale feature segmentation ability is enhanced through the dua-l path attention mechanism and the feature fusion module. The experimental results show that, compared with the original DeepLabV3+ model, the average cross-union ratio (mIoU) of FAD-Net proposed in this paper on the LoveDA dataset has increased by 4.34 percentage points. Through the design of a lightweight network structure, the parameter count is only 4.162× 106 compared to the original DeepLabV3+ model, verifying that the algorithm significantly improves computational efficiency while maintaining semantic segmentation accuracy. |
| Keywords: :land cover classification lightweighting DeepLabV3+ semantic segmentation |
