%0 Journal Article
%T 基于改进DeepLabv3+的高分辨率遥感影像屋顶提取方法<br>Roof Extraction Method of High-Resolution Remote Sensing Image Based on Improved DeepLabv3+
%A 丁澎涛
%A 朱习军
%J Artificial Intelligence and Robotics Research
%P 62-68
%@ 2326-3423
%D 2023
%I Hans Publishing
%R 10.12677/AIRR.2023.122009
%X DeepLabv3+网络能够有效地解决高分辨率遥感图像语义分割的挑战。经过ResNet50骨干网络的支持，我们对DeepLabv3+模型进行了深入的研究，利用Adam梯度下降法和RELU激活函数，有效地处理了遥感影像中的建筑屋顶，提高了语义分割的精度和速度，能够更快地收敛到最优解。同时，空洞空间金字塔池化模块(ASPP)与解码器模块(decoder)的普通卷积部分被并行加权的空洞卷积代替，从而减少参数数量，提升模型的性能。IoU的准确度达到89.2%，超过DeepLabv3+算法，降低了特征提取的误差，同时也大大减少了细节信息的丢失，为最终的语义分割带来了显著的改善。<br />
DeepLabv3+network can effectively solve the challenge of semantic segmentation of high-resolution remote sensing images. With the support of the ResNet50 backbone network, we have conducted in-depth research on the DeepLabv3+ model. Using the Adam gradient descent method and RELU activation function, we have effectively processed the building roof in remote sensing images, improved the accuracy and speed of semantic segmentation, and can quickly converge to the optimal solution. At the same time, the common convolution part of the hole space pyramid pooling module (ASPP) and the decoder module (decoder) are parallelly weighted to reduce the number of parameters and improve the performance of the model. The accuracy of IoU reaches 89.2%, which exceeds the DeepLabv3+ algorithm, reduces the error of feature extraction, and also greatly reduces the loss of detail information, bringing significant improvement to the final semantic segmentation.
%K 遥感图像，DeepLabv3+，语义分割，特征融合，通道注意力<br>Remote Sensing Image
%K DeepLabv3+
%K Semantic Segmentation
%K Feature Fusion
%K Channel Attention
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=65368