单头自注意力和频域-空域融合的水下目标检测

打开文本图片集

关键词：水下目标检测；自注意力机制;Haar小波变换;小目标检测

中图分类号：TP394.1;TH691.9 文献标识码：Adoi：10.37188/OPE.20263404.0671 CSTR：32169.14.OPE.20263404.0671

Underwater object detection based on single-head self-attention and frequency-domain & spatial-domain fusion

LI Dahai，LIAO Jiawei*，WANG Zhendong （School of Information Engineering， Jiangxi University of Science and Technology， Ganzhou 341000，China） * Corresponding author，E-mail： 6720230861@mail. jxust. edu. cn

Abstract： Water refraction，scattering，and uneven ilumination blur target textures. Aquatic organisms are mostly small，camouflaged，and dense. Resource-constrained underwater platforms demand lightweight，real-time models. These factors collectively exacerbate the difficulty of underwater object detection.Therefore，this paper proposed an improved YOLOv8n model based on single-head self-attention and frequency-domain 8. spatial-domain fusion，named YOLOv8n-SD. First，a backbone network enhanced by local-global feature fusion was designed. It used a single-head self-atention mechanism combined with dynamic channel ratio division to efficiently acquire long-range global information from partial channels，and further fused local detail information of efficient feature extraction blocks to realize complementary enhancement of local and global features. Second，a neck network with efficient frequency-domain and spatial-domain fusion was constructed，and a downsampling module using Haar wavelet transform and space-to-depth transform was designed to fuse important high-frequency and spatial information of shallow high-resolution features.At the same time，a fast normalized weighting strategy was adopted to dynamically optimize the efciency of multi-scale feature fusion. On the public underwater datasets URPC2020 and RUOD，the mAP0.5：0.95 and mAP50 metrics of YOLOv8n-SD reach 51.2% ， 85.7% and 50.6% ， 85.0% respectively. Meanwhile，compared with the baseline，the number of parameters is reduced by 42.3% and the computational load is decreased by 17.2% . Comparative experiments further verify that the proposed model exhibits good detection accuracy and robustness in various complex underwater scenarios.

Key words： underwater object detection；self-atention mechanism；Haar wavelet transform； small object detection

1引言

近年来，随着海洋探索在科学研究与经济发展中的战略地位日益凸显，传统人工水下作业“效率低、成本高、风险大”的弊端愈发突出，水下目标检测技术由此逐渐成为主流，在海洋生物保护、海洋资源勘探、水产养殖等场景中得到广泛应用[1-2]。（剩余20355字）