ABSTRACT

Water-body image segmentation is a fundamental operation of many important applications, such as water resources allocation, ecological assessment, flood control, etc. Mainstream neural network based segmentation algorithms are still far from satisfactory for segmenting water-body; complex natural land and water boundaries can easily lead to inaccurate classification with little boundary detail preserved. To improve the performance of water-body segmentation, we propose a novel technique based on feature pyramid enhancement and pixel pair matching. By constructing feature enhancement sub-nets for different scales and superimposing the feature maps together, our technique preserves and transmits more spatial information to the backbone network, hence alleviating the common problem of detail loss in deepened network. Moreover, for each pixel, our technique employs a novel loss term to make the network learn from the classification results of similar neighbouring pixels in order to smooth out small local errors. Experiments on a new water-body dataset, namely DT-1, demonstrate that the proposed methods have improved by at least 1.24% in segmentation precision in comparison with state-of-the-art methods, including the fully connected network (FCN8S), U-Net, SegNet, RefineNet, and DeepLabv3+, which effectively captures the details of water and reduces pixel classification error of the water boundary.

摘要

水体图像分割是水资源分配，生态评估，防洪等许多重要应用的基本操作。复杂的自然土地和水域边界很容易导致分类不准确，而保留的边界细节很少。为了提高水体分割的性能，我们提出了一种基于特征金字塔增强和像素对匹配的新技术。通过构建不同比例的特征增强子网并将特征图叠加在一起，我们的技术可以保留更多空间信息并将其传输到骨干网络，从而缓解了加深网络中常见的细节丢失问题。而且，对于每个像素，我们的技术采用一种新颖的损耗项来使网络从相似的相邻像素的分类结果中学习，以消除较小的局部误差。在新的水体数据集DT-1上进行的实验表明，与包括完全连接网络（FCN8S）的最新技术相比，该方法的分割精度至少提高了1.24％， U-Net，SegNet，RefineNet和DeepLabv3 +，可有效捕获水的细节并减少水边界的像素分类误差。