基于差异增强和双注意力Transformer的遥感图像变化检测

Remote Sensing Image Change Detection Based on Difference Enhancement and Dual Attention Transformer

由于遥感场景中物体的复杂性,光照变化和配准误差都会影响不同时间拍摄的2个图像中目标的变化,探索不同像素之间的关系和更强大识别能力的卷积神经网络可以提高双时相遥感图像变化检测的性能。提出一个基于差异增强的和双注意力机制的Transformer神经网络模型,在孪生网络架构中的特征提取部分引入ResNeXt单元,在不增加参数复杂度的前提下提高准确率;将分层结构的Transformer编码-解码器与通道和空间双注意力模块相结合,获得更大的感受野和更强的上下文塑造能力;该网络还关注双时相图像的差异化特征,通过引入差异增强模块对每个像素进行加权,选择性地对特征进行聚合,最终生成具有高精度的遥感图像变化特征图。通过在变化检测基准数据集LEVIR-CD和DSIFN上进行实验,所提方法对不同建筑物、道路和植被变化情况的检测效果有很大提升,与现有检测模型相比,该方法在F1、IoU和OA这3个评价指标上均好于最好结果。

Due to the complexity of objects in the remote sensing scene,illumination changes and registration errors will affect the changes of the object in two images taken at different time.Exploring the relationship between different pixels and convolutional neural networks with more powerful recognition capabilities can improve bi-temporal performance of change detection in remote sensing images.A novel Transformer neural network model is proposed based on differential enhancement and multi-attention mechanism.The ResNeXt unit is introduced into the feature extraction part of the siamese network architecture to improve the accuracy without increasing the complexity of parameters.The transformer encoder-decoder with hierarchical structure is combined with channel and spatial dual attention modules to obtain a larger receptive field and stronger context shaping ability.In addition,the network also pays attention to the differentiated features of bi-temporal images,weights each pixel by introducing a difference enhancement module,selectively aggregates the features,and finally generates a high-precision remote sensing image change feature map.Through experiments on the change detection benchmark dataset LEVIR-CD and DSIFN,it’s shown that the detection effect of different buildings,roads and vegetation changes is greatly improved.Compared with existing models,this method is better than the best results in F1,IoU and OA.