基于多尺度特征和注意力机制的航空图像分割

Segmentation of aerial image with multi-scale feature and attention model

中文关键词  航空图像分割  建筑像素标记  全卷积神经网络  注意力机制  多尺度特征

英文关键词  aerial image segmentation  build marking  fully convolution neural networks  attention mechanism  multiscale feature

中文摘要

      利用神经网络能通过进行建筑像素标记实现航空图像分割, 但也存在分割边界模糊的问题, 导致分割结果 不理想. 为此, 本文以卷积神经网络U–net和FCN–8s基本网络模型, 实现端到端训练. 在此基础上, 结合建立了全卷 积神经网络结合多尺度特征和注意力机制的网络模型, 提升了分割边界的清晰度. 将多尺度特征和注意力机制的模 型与基本模型进行对比, 分析了真实与预测之间的相关度和相似度, 并将预测结果进行对比. 实验结果表明, 结合 多尺度特征和注意力机制的分割模型, 分割边界更清晰, 相对于相同训练规模的全卷积网络交并比高2%, Dice系数 高3%, 得到较好的分割结果.

英文摘要

      Employing neural network to automatically segment aerial image by marking building pixels. However, there is also the problem of segmentation boundary blurring, resulting in the segmentation is not ideal. In this paper, the fully convolutional neural network (FCNs), U–Net and FCN–8s, are employed as basic model to train end-to-end. And then, a network model combining the multi-scale feature and attention mechanism is established, and the clarity of the segmentation boundary is improved. The model of multi-scale feature and attention mechanism is compared with basic model, and the relativity and similarity between the prediction and ground truth (GT) are analyzed, and the prediction results are compared. The results show that, when the model of combining multi-scale features and attention mechanism is adopted to aerial images, segmentation boundary is clearer and the boundary detail processing is better. Compared to the original full convolutional neural networks of same training scale, the Intersection over Union (IoU) is 2% higher and the Dice coefficient is 3% higher for model that combine multi-scale features and attention mechanism, and a better segmentation result is obtained.