Study on Underwater Image Semantic Segmentation Based on UNet

The study of the underwater world has a profound impact on human society, with research on marine life playing a significant role. Among many classical image segmentation network models, the UNet network model exhibits excellent performance on small sample datasets. Due to this characteristic, it has been widely applied in the medical field. However, this property also makes it suitable for underwater images with limited data availability. Moreover, to deploy segmentation network models on underwater mobile devices, it is necessary to significantly reduce the number of model parameters and improve computational efficiency. Therefore, this paper proposes an improved model based on UNet, which integrates Convolutional layers and Multi-Layer Perceptrons (MLP) to optimize UNet, reducing the number of parameters by more than 20 times. Although research on semantic segmentation of underwater images continues to advance, many studies have yet to fully consider the complexity of the underwater environment and the camouflage characteristics of marine organisms. This paper introduces a novel data augmentation method, which applies Fourier transform to alter the low-frequency components of an image, generating new images that retain the same semantic content as the original but with different styles. A large number of experimental results demonstrate that the model proposed in this paper reduces the number of parameters by more than 20 times while improving the mean Intersection over Union (mIoU) by 11%.