Second-place Solution of 5th GaoFen Challenge of Sea Ice Segmentation Channel

It can be observed from the training data that the structures of sea ices are complex and the details of contours should be well preserved. Therefore, we modify the widely exploited U-Net network by replacing the encoder part with pre-trained networks via the ImageNet dataset and re-design decoders. For the encoder, we remove the fifth stage and just remain the first four stages. Since the feature maps of the fifth stage generally capture high-level semantics of the input images, they cannot be of great help for localizing the boundary structures of sea ices at pixel-level. In addition, we exploit the deep supervision strategy¹ by involving two more segmentation heads from the last two stages of decoders, in order to force the low-level features better capturing the contours of sea ices. For decoders, we would like to achieve a relatively lightweight CNN block to increase the inference speed and adopt the one proposed in [2], which exploits the residual learning scheme with the depthwise and pointwise convolutions involved.