Underwater Object Detection with Swin Transformer

Abstract

Underwater target detection algorithms are of great significance in exploration of the oceans, as well as search and rescue operations. The photographs that were taken underwater are very distinct from those that were taken on land because of the effect that water has on the way light travels through it. Images captured underwater are frequently distinguished by a color palette that ranges from blue to green, as well as by varying degrees of distortion and blurriness. In addition, marine organisms frequently have distinctive ways of life, which presents significant difficulties when attempting to locate objects sub-merged below the surface. The application of Natural Language Processing (NLP) Transformer to the visual field has become increasingly popular as of late. In particular, the use of the straightforward and efficient Vision Transformer (ViT) enables us to see the potential of Transformer in the visual field. On the foundation of the conventional Transformer, Swin Transformer takes the concept of feature extraction from a convolutional neural network and applies it to the task of obtaining multi-scale features of images in a hierarchical fashion. It is possible to employ it as a generic backbone network to overturn the irreplaceable status of convolutional networks in the visual field, where it has a good influence on all downstream tasks in the visual field. As a result, we attempt to use Swin Transformer in the field of underwater object detection, and the results of our experiments indicate that Swin-B has a similar detection effect to Cascade R-CNN with ResNeXt101 backbone.