基于全卷积网络提出了一种图像分割模型以获取目标分割结果,模型包含两个结构相同的深层神经网络分支,每个分支采用卷积-反卷积的结构实现特征提取和从特征恢复目标区域;两个分支接收不同类型图像输入,将来源于两个分支的结果通过加权融合得到最终的分割结果。模型融合了不同图像源的多级尺度特征,在训练样本数有限的情况下,通过数据增强使训练得到的模型稳健性更强。在光学图像数据集Weizmann horse和遥感影像数据集Vaihigen上进行实验,并与相关文献进行比较,结果表明,所提模型具有更高的目标分割完整度和最优的分割性能,在训练数据有限、形态各异、尺度变化较大等的遥感影像建筑物提取中取得了理想的结果,表明该模型可应用于复杂的遥感影像目标分割。

Based on the full convolutional network, an image segmentation model is proposed to obtain the target segmentation results. This model consists of two deep neural network branches with the same structures. As for each branch, a convolution-deconvolution structure is adopted to implement the feature extraction and to recover the target area from the features. These two branches receive different image inputs, and then the final segmentation results are obtained via the weighted fusion of the results from these two branches. This model combines the multi-level scale features of different image sources, and the training model is more robust through data enhancement when the number of training samples is limited. The experiments are carried out on the optical image dataset of Weizmann horse and the remote sensing image dataset of Vaihigen. The comparison with the related literatures is also made. The results show that the proposed model has a higher target segmentation integrity and an optimal segmentation performance. The ideal extraction results of remote sensing image buildings under the conditions of limited training data, various shapes, large scale changes and so on, indicate that the proposed model can be applied to the complex remote sensing image object segmentation.