Small-Object Semantic Segmentation of Satellite Ship Images Using Modified U-Net With Morphological Loss

Satellite aerial images have been successfully used for many applications. However, it is still challenging to detect small objects due to low-resolution and environmental conditions. To overcome this challenge, we propose segmentation methods using atrous spatial pyramid pooling (ASPP) and modified residual U-Nets with weighted loss functions. ASPP is a spatial pyramid pooling method with different atrous convolutions that can enlarge receptive fields without reducing the feature map resolution. A residual block enables the network to optimize more easily during training with skip connections. We adapted this residual block to the U-Net and modified it with atrous convolutions to enlarge the receptive field. We also applied a loss function computed by combining a focal loss function and a new loss function for small objects. Furthermore, the loss function for small objects is computed only for small objects by applying white top-hat processing. We tested the proposed method using the Airbus ship segmentation dataset. The experimental results show that the proposed methods produced improved Dice and mIoU performance by 2 to 8 percent compared to existing segmentation models. The proposed methods successfully detected some small objects missed by the existing segmentation models.