渐进自适应特征融合的轻量化火焰检测算法研究  

Study on Lightweight Flame Detection Algorithm with Progressive Adaptive Feature Fusion

作　　者：李啸澜 马勇[1] LI Xiaolan;MA Yong(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机科学》2025年第4期64-73,共10页Computer Science

基　　金：国家自然科学基金(61773210)。

摘　　要：针对视觉安防系统在边缘计算平台部署火焰检测模型时面临的精度与实时性难以平衡的问题,提出一种渐进自适应特征融合的轻量化火焰检测算法。首先,设计轻量级稀疏卷积算子降低模型计算复杂度与内存访问开销。其次,针对分组卷积的通道间信息交互缺陷,基于残差思想构建长距离上下文特征增强的轻量级特征提取组件。为解决深度骨干网络中特征丢失及背景干扰问题,创新性地提出基于高频增强的轻量级特征强化机制,优化空间域和通道域参数,缓解背景干扰问题。在此基础上,建立特征增强-渐进自适应特征融合框架,促进不同尺度特征图充分融合,提高特征图利用率,增强对多尺度目标的识别效果。实验结果表明,所提方法在实时推理速度最高达到27.1 FPS的同时,参数量降低至2.1×10^(6),较基准模型减少69.5%,并达到83.4%的mAP@0.5检测精度,显著优于现有主流方法。In response to the challenge of balancing accuracy and real-time performance when deploying flame detection models on edge computing platforms for visual security systems,a lightweight flame detection algorithm featuring progressive adaptive feature fusion is proposed.Firstly,a lightweight sparse convolution operator is designed to reduce the model’s computational complexity and memory access cost.Subsequently,to address the shortcomings of inter-channel information exchange in grouped convolutions,a lightweight feature extraction component is constructed based on the residual concept,enhancing long-distance contextual features.To tackle the issues of feature loss and background interference in deep backbone networks,an innovative lightweight feature enhancement mechanism based on high-frequency augmentation is proposed,optimizing the parameters in both spatial and channel domains to mitigate background disturbances.On this basis,a feature enhancement-progressive adaptive feature fusion framework is established to facilitate the thorough integration of feature maps at different scales,thereby improving the utilization of feature maps and enhancing the recognition effectiveness of multi-scale targets.Experimental results demonstrate that this method achieves a real-time inference speed of up to 27.1 FPS,reduces the parameter count to 2.1 M,which is a 69.5%reduction compared to the baseline model,and attains a detection accuracy of 83.4%mAP@0.5,significantly outperforming existing mainstream methods.

关 键 词：深度学习 计算机视觉 目标检测 轻量化神经网络 特征提取网络 特征融合网络 特征增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]