VVNet:Underwater Object Detection Network Based on Vision Transformer and Vision Retnet for Underwater Robot Picking

Abstract

Underwater object detection is an important task in computer vision, but it is usually difficult due to the complex underwater environment. In recent years, underwater object detection methods based on deep learning have been proposed and have achieved promising results. However, there still exist issues with weaker extraction of global features from images. To tackle this issue, we propose a novel neural network, VVNet, which incorporates the newly introduced RMTBFA and CVit modules, achieving superior performance in extracting global features. RMTBFA integrates the core module of the Vision Retinet model, the Manhattan Attention Mechanism, while CVit incorporates the Vision Transformer module, which facilitates the extraction of global features from images. Furthermore, we introduce Shape-IoU as a loss function for IoU, which further enhances the accuracy of our model. Finally, we tested on the two datasets of URPC2017 and URPC2018. The results showed that our network model improved the underwater object detection task indicators.