Underwater Target Detection Based on Enhanced YOLOv9 Model

Abstract

Underwater target detection is crucial for ocean exploration, yet Traditional ground target detection method struggle under harsh underwater conditions. Issues such as inconsistent lighting, low contrast, and physical obstacles can obscure targets, making it particularly difficult for conventional approaches to accurately detect low-resolution images and small targets, which are further compromised by ambient noise and variable target sizes. These challenges impede accurate feature extraction and slow the detection process. Consequently, advanced detection technologies are imperative to overcome these obstacles and drive progress in marine environmental engineering. In light of the aforementioned challenges, we present an improved iteration of YOLOv9, named the enhanced YOLOv9 model, tailored to bolster detection precision for images of low resolution and minute targets. This upgraded model provides the capacity to fine-tune parameters and modify the shapes of convolutional kernels, leading to a decrease in the computational burden on convolutional neural networks (CNNs). Such optimization notably enhances the efficacy of target detection in complex underwater settings, effectively tackling inherent obstacles and fortifying the model's overall resilience. Ablative experiments on Underwater Robot Picking Contest (URPC) and Underwater Object Detection (UOD) datasets and comparative analysis of current mainstream detectors show that this model demonstrates clear superiority in both detection capability and efficiency.