An efficient approach to detect and segment underwater images using Swin Transformer

Abstract

In managing the fish environment and underwater ecosystem, there's been a big shift from old-generational schoolings, hands-on methods to high-technologies, automatic ways of doing things, because of some clever computer technology, using a type of Artificial Intelligence smartness known as Convolutional Neural Networks, or CNNs in short. These smart systems help deal with the tricky task of keeping tabs on where fish stay out in deeper underwater, particularly when it's hard to see underwater and when the fish are all over the place. There are some pretty good tools and techniques already out there, like DPANet and a few others with names like PIFS-Net, EFS-Net, and MFAS-Net, which have been helping a bunch of researches and developments for underwater ecosystem. But Deeper water environment still needs a newer and better way to get the tasks done. So, to answer this major question, this hybrid architecture rolls out a new combination that takes the best bits of transformer architecture known as Swin Transformer and another mechanism known as ConvMixer, making it in a form of SwinUNet Architecture. This Swin Transformer is well defined at picking up on the spatial information from underwater deeper images in our case, which is a key point to be highlighted in really getting the fish's niche turf and enquiring their ecosystem. And when the underwater turns misty, the specialized implementation of ConvMixer modules incorporated with Swin Transformer increases its ability to sort out the fish from the rest of the underwater world. And this makes it to perform even more better compared with the existing state of arts model, even under Modern Underwater Machine Vision System conditions. This model even over takes to utilization of Few-Shot Learning model for training it in this lesser Underwater Images, with its capability of identifying the patterns and differentiations in both larger and smaller class objects in the given ground truth image. In this new approach, the paper combines the smart focusing methods with a way to blend different signals to better mIoU score, making it super useful for having a deeper understanding on fish-ecosystem and also their underwater ecosystem. While comparing it with the pre-existing state of arts model and the booming Image Segmentation models (YOLO v8, Pix 2 Pix GAN model, Auto Encoders, UNet), the hybrid approach outperformed the existing model results on this Semantic Segmentation for Underwater Imagery (SUIM) dataset, which is been disclosed in the further sections below.