融合余弦退火与空洞卷积的遥感影像语义分割

Remote sensing image semantic segmentation method combining cosine annealing with atrous convolution

为了捕捉遥感影像中丰富的上下文信息与多尺度的地物信息,改进集成模型的策略,提高语义分割精度,提出一种融合周期递增余弦退火与多尺度空洞卷积的高分辨率遥感影像语义分割方法。方法引入多尺度并行的空洞卷积,有利于捕捉更大范围的上下文信息,在不增加参数的情况下,提高网络对多尺度对象的辨识能力;使用全连接条件随机场引入空间和边缘的上下文信息,提高网络对遥感影像的细节分割能力;引入周期递增的余弦退火策略调整学习率,获得合适数量的局部最优解,集成局部最优解进一步提升网络在像素上的分类能力。在Gaofen Image Dataset数据集上的实验结果表明,多尺度并行空洞卷积可以充分捕捉遥感影像上的多尺度地物信息,能有效辨识复杂对象;空间和边缘上下文信息的引入使语义分割对象的边界辨识更精准;周期递增余弦退火策略能明显减少集成模型的推理时间,模型的总体精度与Kappa系数均优于目前主流的语义分割模型。

This study aims to capture the rich context information and multiscale feature information in remote sensing images,improve the integrated model strategy,and enhance the accuracy of semantic segmentation.Thus,this study proposes a high-resolution remote sensing image semantic segmentation method using cosine annealing with increasing period and multiscale atrous convolution.The multiscale parallel atrous convolution helps the network capture context information in a larger range and improves the ability of the network to recognize multiscale objects without increasing parameters.The method in this study uses the atrous convolution while discarding the pooling operation to maintain the spatial resolution.Meanwhile,the method adopts the fully connected conditional random field to add spatial and edge context information for making up for part of the position information missed by the atrous convolution.As a result,the outline of extraction objects by semantic segmentation fits the ground truth better.Moreover,the cosine annealing strategy with increasing period is introduced to adjust the learning rate and obtain a suitable number of local optimal solutions.We integrate the local optimal solutions in the method to further improve the pixel classification ability of the network.The overall accuracy and kappa coefficient of the proposed model,which are 86.6%and 81.8%,respectively,are better than those of the current advanced semantic segmentation models.The experimental results performed on the Gaofen image dataset show that the fusion of image context information and multiscale feature information can effectively identify objects with complex structures.Moreover,the model coupled with the period-increasing cosine annealing strategy could obtain better semantic segmentation accuracy than and less inference time than that coupled with the equal-period cosine annealing strategy.