HKB-YOLO: transmission line fire detection method based on hierarchical feature fusion

作     者：Ying, Zhilei Meng, Yanan Chen, Ruoxi Lou, Jianlou 

作者机构：School of Information and Control Engineering Jilin Institute of Chemical Technology Jilin 132022 China Department of Automation Jilin Institute of Chemical Technology Jilin 132022 China School of Computer Engineering Northeast Electric Power University Jilin 132012 China 

出 版 物：《International Journal of Security and Networks》 (Int. J. Secur. Netw.)

年 卷 期：2024年第19卷第4期

页      面：188-198页

核心收录：

学科分类：0808[工学-电气工程] 080802[工学-电力系统及其自动化] 0817[工学-化学工程与技术] 08[工学] 0807[工学-动力工程及工程热物理] 0837[工学-安全科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：We thank the anonymous reviewers for their constructive feedback and the members involved for their support and assistance 

主　　题：Premixed flames 

摘      要：Timely transmission line fire inspections are vital for power system safety. Although deep learning models are widely used for flame detection, struggle with small target recognition due to background interference and are vulnerable to input data perturbations, posing security risks. This study introduces a hierarchical feature fusion method based on the you only look once Version 8 (YOLOV8) framework. It employs high-performance GPU network version 2 (HGNetV2) to enhance small-target feature extraction while reducing computational complexity. A spatial pyramid pooling-fast module with large separable kernel attention is designed to highlight key features and suppress background noise. The proposed bidirectional slim neck structure (BiSlimneck) reduces feature loss. Adversarial training enhances the model’s robustness. Experimental results show a 1.4% accuracy improvement, with a reduction in parameters and complexity by 18.3% and 21.6%, respectively. After adversarial training, the accuracy of the network in the face of attacks increased by 10% compared to the pre-training. Copyright © 2024 Inderscience Enterprises Ltd.