计算机科学
稳健性(进化)
计算
人工智能
数据挖掘
机器学习
分布式计算
模式识别(心理学)
算法
生物化学
化学
基因
作者
Jiguo Yu,Hao Zheng,Longhan Xie,Lei Zhang,Mei Yu,Jiqing Han
标识
DOI:10.3389/fnbot.2023.1315251
摘要
Unmanned surface vessel (USV) target detection algorithms often face challenges such as misdetection and omission of small targets due to significant variations in target scales and susceptibility to interference from complex environments. To address these issues, we propose a small target enhanced YOLOv7 (STE-YOLO) approach. Firstly, we introduce a specialized detection branch designed to identify tiny targets. This enhancement aims to improve the multi-scale target detection capabilities and address difficulties in recognizing targets of different sizes. Secondly, we present the lite visual center (LVC) module, which effectively fuses data from different levels to give more attention to small targets. Additionally, we integrate the lite efficient layer aggregation networks (L-ELAN) into the backbone network to reduce redundant computations and enhance computational efficiency. Lastly, we use Wise-IOU to optimize the loss function definition, thereby improving the model robustness by dynamically optimizing gradient contributions from samples of varying quality. We conducted experiments on the WSODD dataset and the FIOW-Img dataset. The results on the comprehensive WSODD dataset demonstrate that STE-YOLO, when compared to YOLOv7, reduces network parameters by 14% while improving AP50 and APs scores by 2.1% and 1.6%, respectively. Furthermore, when compared to five other leading target detection algorithms, STE-YOLO demonstrates superior accuracy and efficiency.
科研通智能强力驱动
Strongly Powered by AbleSci AI