| 引用本文： |
-
陆建波,彭俊桂,霍雷刚,刘晓彬.一种轻量级DeepLabV3+遥感图像分割方法[J].广西科学,2025,32(2):374-385. [点击复制]
- LU Jianbo,PENG Jungui,HUO Leigang,LIU Xiaobin.A Lightweight DeepLabV3+ Model for Remote Sensing Image Segmentation[J].Guangxi Sciences,2025,32(2):374-385. [点击复制]
|
|
| 摘要: |
| 针对遥感图像语义分割中的物体边界分割不全、模型参数量大和占用内存多等问题，本研究提出一种轻量级DeepLabV3+遥感图像分割方法(L-DeepLabV3+)。在模型参数量更小的情况下，该方法能够提升物体边界分割精度。具体而言，L-DeepLabV3+在残差模块中采用维度下降策略，通过减少输出特征图的通道数，降低模型参数量，将MobileNetV2主干网络中的倒残差模块替换为提出的降维残差模块，重构特征提取网络；为了加快模型训练速度，该方法在DeepLabV3+中的空洞空间金字塔池化层(Atrous Spatial Pyramid Pooling,ASPP)中使用深度可分离卷积(Depthwise Separable Convolution,DSConv)；此外，为解决样本类别不均衡问题并提高分割准确率，将交叉熵损失函数(Cross-Entropy loss,CE loss)和Dice loss合并为新的损失函数进行训练；最后，通过将归一化层与卷积层融合、将1×1卷积和恒等残差转换成3×3卷积实现模型参数重构，从而在推理时实现模型轻量化。在DLRSD、WHDLD、UDD6等数据集上的实验结果表明，改进的L-DeepLabV3+模型的参数量仅3.5 M，有较高的分割准确率和训练效率，画面每秒传输帧数(Frames Per Second，FPS)可达到90.2。 |
| 关键词: 遥感图像 DeepLabV3+ 深度可分离卷积 轻量化 |
| DOI：10.13656/j.cnki.gxkx.20250624.018 |
| 投稿时间：2023-05-11修订日期：2023-09-25 |
| 基金项目:广西重点研发计划项目(桂科AB21076021)资助。 |
|
| A Lightweight DeepLabV3+ Model for Remote Sensing Image Segmentation |
|
LU Jianbo1,2, PENG Jungui2, HUO Leigang1,2, LIU Xiaobin2
|
| (1.Guangxi Key Laboratory of Human-machine Interaction and Intelligent Decision, Nanning Normal University, Nanning, Guangxi, 530100, China;2.School of Computer and Information Engineering, Nanning Normal University, Nanning, Guangxi, 530100, China) |
| Abstract: |
| Aiming at the incomplete object boundary segmentation,large number of model parameters,and large memory consumption in semantic segmentation of remote sensing images,this study proposes a lightweight DeepLabV3+ model (L-DeepLabV3+) for remote sensing image segmentation.This model can improve the accuracy of object boundary segmentation with a smaller number of parameters.Specifically,this model adopts a dimensionality reduction strategy in the residual module to reduce the number of model parameters by decreasing the number of channels in the output feature map,and reconstructs the feature extraction network by replacing the inverted residual module in the MobileNetV2 backbone network with the proposed dimensionality-decreasing residual module.To accelerate the training speed,the model uses Depthwise Separable Convolution (DSConv) in the Atrous Spatial Pyramid Pooling (ASPP) of DeepLabV3+.In addition,to solve the sample category imbalance and improve the segmentation accuracy,Cross-Entropy loss (CE loss) and Dice loss are combined into a new loss function for training.Finally,model parameter reconstruction is achieved by fusing the normalization layer with the convolutional layer,and converting the 1×1 convolution and constant residuals into 3×3 convolution,for model lightweighting in inference.The experimental results on the DLRSD,WHDLD,and UDD6 datasets show that the improved L-DeepLabV3+ model has high segmentation accuracy and training efficiency,with a model parameter number of only 3.5 M.Meanwhile,the number of transmitted Frames Per Second (FPS) of the screen can reach 90.2. |
| Key words: remote sensing images DeepLabV3+ depthwise separable convolution lightweight |
