Underwater Small Target Detection Based on Improved YOLOv7 Model

Abstract

In the domain of marine environmental engineering, rapid and precise detection of underwater targets holds substantial significance. However, extant technologies for object detection are plagued by diminished accuracy and protracted processing speed when tasked with identifying occluded and diminutive targets in aquatic settings. In response to these impediments, this article has made innovative improvements on the basis of YOLOv7 to improve these issues. Firstly, we designed a G-ELAN module, which convolves low-level features and multiplies them element by element with high-level features, achieving multi-level feature fusion and effectively reducing target occlusion interference caused by underwater plankton. Secondly, the Residual-FasterNet module was proposed, which integrates residual strategies based on the T-shaped receptive field of FasterNet. This amalgamation enhances the model's ability to capture small target features. Evaluations conducted on the URPC2021 public dataset and the Zhanjiang competition dataset corroborate that our YOLOv7-GR model surpasses prevailing advanced methodologies in performance.The source code will be made publicly available at https://github.com/zzuilcc/YOLOv7-GR.git}{https://github.com/zzuilcc/YOLOv7-GR.git.