基于轻量化的神经网络火焰检测算法研究  

Research on Flame Detection Algorithm Based on Lightweight Neural Network

作　　者：陈义鑫 王婷[1] 杨万扣[2] CHEN Yi-xin;WANG Ting;YANG Wan-kou(School of Information Science and Technology,Nanjing Forestry University,Nanjing 210037,China;School of Automation,Southeast University,Nanjing 211189,China)

机构地区：[1]南京林业大学信息科学技术学院,江苏南京210037 [2]东南大学自动化学院,江苏南京211189

出　　处：《计算机技术与发展》2024年第11期200-206,共7页Computer Technology and Development

基　　金：国家自然科学基金(62276061,62006041);江苏省研究生科研与实践创新计划项目(SJCX21_0338)。

摘　　要：为了改善复杂背景情况下火焰检测算法检测效果较差、对小目标不敏感和计算量过大等缺陷,该文设计出一种轻量化的目标检测算法——GS-YOLOv5s。研究将鬼影混洗卷积瓶颈模块(GS bottleneck)应用于特征提取网络,提出跨阶段特征提取网络——GS-C2,通过对局部区域进行特征分流,使模型能够更专注地学习目标周围的局部特征,从而提高复杂背景下的目标检测精度;然后在模型的颈部网络中使用轻量化卷积GSConv,在实现模型轻量化的同时提高检测精度;最后通过融入选择性注意力LSK模块扩大感受野,更好地捕捉到全局上下文信息,提供更全面的场景理解,使网络更好地理解和响应小目标。数据集测试结果表明,与YOLOv5s基准模型相比,该算法结构整体参数量和浮点运算量分别减少了约20%和21%,同时mAP_(0.5)相较于YOLOv5s提高了4.6百分点。实验结果表明,GS-YOLOv5s既提高了检测精度,又满足了轻量化以及实时检测的需求,极大提高了火焰检测算法的实用性。In order to improve the defects of flame detection algorithms in complex background,such as poor detection effect,insensitivity to small targets and excessive computation,a lightweight object detection algorithm,GS-YOLOv5s,is designed.GS-C2,a bottleneck module,is applied to a feature extraction network,and a cross-stage feature extraction network is proposed.GS-C2 makes the model more focused on learning local features around targets,thus improving the accuracy of target detection in complex backgrounds.Then,the lightweight convolutional GSConv is used in the neck network of the model to improve the detection accuracy while realizing the lightweight model.Finally,by integrating selective attention LSK module to expand the receptive field,better capture the global context information,provide a more comprehensive understanding of the scene,so that the network can better understand and respond to small targets.Data set test results show that compared with the YOLOv5s benchmark model,the overall structure parameter number and floating point computation amount of the proposed algorithm are reduced by about 20% and 21% respectively,while mAP_(0.5 )is increased by 4.6 percentage points compared with YOLOv5s.The experimental results show that GS-YOLOv5s not only improves the detection accuracy,but also meets the requirements of lightweight and real-time detection,which greatly improves the practicability of flame detection algorithm.

关 键 词：目标检测 GSConv 跨阶段局部网络 注意力机制 SK模块 

分 类 号：TP249[自动化与计算机技术—检测技术与自动化装置]