A New Network Structure for Semantic Segmentation of Ship Targets in Remote Sensing

Accurate detection of ship targets is a research hotspot in computer vision. Most of the researches have achieved instance-level detection in the way of bounding box. But we intend to achieve more accurate detection of ship targets in pixel-level through semantic segmentation. However, there are still two main challenges: the first one is the difficulty to segment small targets caused by the difference among ship targets' scales, and the other one is the lack of localization information caused by insufficient recovery ability of the decoder part. In this paper, we propose an effective solution. First, a multi-scale pooling fusion module is proposed to fuse multi-scale feature maps and acquire more multi-scale context information, then we improve the capability of precise decoding by taking the place of convolution operation with deconvolution in the decoder part to gather more localization information. At last, we integrate above two schemes into an encoder-decoder symmetry training network with less training parameters and less training time. Furthermore, we construct a dataset for ship semantic segmentation called HRSC2016-SS by labeling HRSC2016 dataset to evaluate our solution. Experiments show that comparing with the existing methods, our proposed solution has a better performance.