深度卷积神经网络支持下的遥感影像语义分割

针对高分遥感影像语义分割面临的类别不平衡和上下文信息利用不充分问题,本文提出了一种优化的DeeplabV3+算法。首先通过修改交叉熵损失函数,解决数据不平衡问题;其次使用Vortex Pooling取代ASPP模块提高上下文信息;然后采用多尺度输入充分利用图像的多尺度信息,并用投票策略进行特征融合提高图像分割准确性;最后使用形态学作后处理消除拼接痕迹和噪声。在CCF大赛的数据集上进行训练,并与其他经典语义分割算法进行比较。试验结果表明,该算法充分利用上下文信息,有效减少了错误分类,且使分割边界更精确,尤其对于线状目标的捕捉能力更强;在整幅测试影像上的MIoU可达85.21%,明显优于SegNet、U-Net算法。

Remote sensing image semantic segmentation supported by deep convolutional neural networks

Aiming at the problem of category imbalance and insufficient utilization of context information in high-score remote sensing image semantic segmentation, this paper proposes an optimized DeeplabV3 + algorithm. Firstly, the data imbalance problem is solved by modifying the cross-entropy loss function. Secondly, it replace the ASPP module with Vortex Pooling to improve the context information. Then it use multi-scale input to make full use of the multi-scale information of the image. And then it use the voting strategy for feature fusion to improve the accuracy of image segmentation. Finally, morphology is used for post-processing to eliminate stitching marks and noise. Train on the CCF contest dataset and compare it with other classic semantic segmentation algorithms. The experimental results show that the algorithm in this paper makes full use of contextual information, effectively reduces misclassification, makes segmentation boundaries more accurate, and captures linear targets more effectively. The MIoU on the entire test image can reach 85.21%, which is significantly better than the SegNet and U-Net algorithms.