摘要
为解决传统DeepLabv3+算法在遥感影像变化检测上出现的边缘目标分割不精确、分类结果差的问题,提出了一种改进DeepLabv3+的高分辨率遥感影像变化检测方法。首先,基于深度分离卷积与空洞卷积构建了DeepLabv3+模型,大大降低了模型的计算量和参数量。其次,通过引入异感受野改进池化金字塔结构,同时在解码器模块中加入多尺度特征张量,对中间流结构进行残差改造,优化Xception骨干网络,并通过设置权重系数对网络通道进行权重配置优化,从而改进DeepLabv3+模型。最后,采用非生成性和生成性样本扩充方法构建数据集,并通过实验对比分析了所提方法的检测精度与泛化性能。实验结果表明,所提方法能够有效改善图形的输出分辨率和细节特征,具有良好的泛化性能和较高的检测准确率,且与其他对比方法相比,所提方法的图像检测准确率较高,整体精度指标最高可达96.4%。
To solve the problem of inaccurate segmentation of edge targets and poor classification results in the traditional DeepLabv3+ algorithm in remote sensing image change detection, an improved DeepLabv3+ highresolution remote sensing image change detection method is proposed. First, a DeepLabv3+ model is developed based on deep separation and hole convolutions, which significantly reduces the amount of calculation and model parameters. Second, the pooling pyramid structure is improved by introducing different receptive fields. Moreover,multiscale feature tensors are added to the decoder module;the intermediate stream structure is reconstructed;and the Xception backbone network is optimized. Then, the network channel is adjusted by setting weight coefficients.The weight configuration is optimized to improve the DeepLabv3+ model. Finally, non-generative and generative sample expansion methods are used to develop the dataset. The detection accuracy and generalization performance of the proposed method are confirmed via experimental comparison and analysis. The experimental results demonstrate that the proposed method can effectively improve the output resolution and detailed characteristics of graphics. This shows that the proposed method has good generalization performance and higher detection accuracy compared to other traditional methods. Furthermore, the proposed method has the highest image detection accuracy compared with other traditional methods, and the overall accuracy index can reach 96. 4%.
作者
常振良
杨小冈
卢瑞涛
庄昊
Chang Zhenliang;Yang Xiaogang;Lu Ruitao;Zhuang Hao(College of Missile Engineering,Rocket Force Engineering University,Xi’an 710025,Shaanxi,China;The 32023 Unit of the People’s Liberation Army,Dalian 116085,Liaoning,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2022年第12期483-494,共12页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61806209)。
