RepUNet:A Fast Image Semantic Segmentation Model Based on Convolutional Reparameterization of Ship Satellite Images

In this paper, a fast image semantic segmentation model based on convolutional reparameterization of ship satellite images is developed. The salient features of the proposed approcach are: (1) Various semantic segmentation networks are evaluated using high-resolution ship satellite data. (2) A high-precision and high-speed semantic segmentation network, termed RepUnet, for detecting ships according to complex characteristics of the actual marine situation is developed. The semantic segmentation network of RepUnet is divided into four parts, namely encoding layer, decoding layer, convolutional reparameterization and batch normalization feature fusion. The RepUNet has the following characteristics: (1) The structure of 33 convolution is used as the basic skeleton, and the basic propagation speed is high. (2) The 11 convolution and identity branch layer are adopted to enhance feature extraction. (3) All the branches of the encoding layer are reparameterized and the parameters are embedded in the convolution kernel of 3*3. The branch structure is abandoned when inferring so that feature extraction is strengthened. Batchnormalization of each branch is fused to make the feedforward speed faster. Validation of the RepUnet demonstrates that it outperforms other semantic segmentation networks. The IoU and feedforword speed (Ffs) of 22.26% and 153.11 %, respectively are significantly higher than those achieved by the UNet.