基于L-DeepLabv3+的轻量化光学遥感图像道路提取

Lightweight optical remote sensing image road extraction based on L-DeepLabv3+

针对DeepLabv3+在进行光学遥感图像道路提取任务时,存在模型参数量大、细节提取效果差等问题,提出一种改进DeepLabv3+的轻量化道路提取模型L-DeepLabv3+。首先通过将主干网络替换为MobileNetv2来减少模型参数量;其次,在编码层中设计一个改进的空洞空间卷积池化金字塔模块。该模块通过嵌入一个通道空间并联注意力模块和YOLOF模块来增强模型特征表达能力,并且将普通卷积替换为深度可分离卷积进一步减少模型参数量;最后组合采用Dice_loss和Focal_loss作为损失函数来解决正负样本不均衡问题。实验结果表明:L-DeepLabv3+在DeepGlobe Road数据集上进行道路提取的交并比达到68.40%,像素准确率达到82.67%,且模型参数量仅为5.63 MB,FPS达到72.3,与其他模型相比具有明显提升,实现了模型精度与轻量化之间更好的平衡。

To address the problems of large number of model parameters and poor detail extraction in DeepLabv3+for optical remote sensing image road extraction task,a light-weight road extraction model L-DeepLabv3+is proposed to improve DeepLabv3+.Firstly,the number of model parameters is reduced by replacing the backbone network with MobileNetv2;secondly,an improved void space convolutional pooling pyramid module is designed in the coding layer.This module enhances the model feature expression capability by embedding a channel space parallel attention module and YOLOF module,and replaces the normal convolution with deep separable convolution to further reduce the number of model parameters;Finally,Dice_loss and Focal_loss are combined as loss functions to solve the positive and negative sample imbalance problem.The experimental results show that L-DeepLabv3+achieves 68.40%intersection ratio and 82.67%pixel accuracy for road extraction on DeepGlobe Road dataset,and the number of model parameters is only 5.63 MB,and the FPS reaches 72.3,which is a significant improvement compared with other models,and achieves a better balance between model accuracy and light weight.