RMP-Net: A high-performance super-resolution underwater scene segmentation network based on parameter reconstruction

Ocean exploration is the important area which focus on humanity strive. Nowadays, autonomous underwater explorer are made by many countries to realize seabed. However, high visual system is the most important technology. This is caused by the complexity of environment and attenuation of light. To solve the problem, in one hand we improvement the construction by tradition model. On the other hand, an algorithm is proposed to perform semantic segmentation of seabed scenes. So we can intensifier Feature extraction with Super-resolution fusion. In complex environments, the use of a structurally reparameterized backbone network allows for a more effective extraction of target features, by combining semantic information from multiple scales through sub-pixel super-resolution, we can attain outstanding segmentation performance for ocean scenes. Multiclass segmentation and dichotomous segmentation experiments were conducted in SUIM. In this research, we can say the effect is the best one even compare with up-to-dated methodology, our indicators reached 84.52% in miou and 92.33% mPA, respectively. Our model and code will be freed in Github.