Scholarlink

<p>Due to variations in wind direction, the shape of fire can undergo significant changes. Traditional convolution, with its fixed weights, cannot adaptively extract image features, making it challenging to handle images with complex backgrounds and unevenly distributed fire of different shapes.Therefore, to address this challenge, we investigated the unique structure of fire and proposed a multi-dimensional dynamic convolutional fire detection method (MDC-WDA) based on wind direction aware. Firstly, a wind direction-aware convolution kernel (WDA) is proposed to adaptively focus on changes in fire and smoke by capturing wind direction, enhancing the model’s perception ability and accurately capturing the features of object structures in images. Secondly, in the feature extraction network CSPDarknet, a multi-dimensional dynamic convolutional network (MDC) is proposed to adaptively change the weights of multiple dimensions of convolution, extract prior knowledge provided by different views, deal with localization difficulties caused by complex scenes and uneven distribution. Finally, utilizing the excellent ability of state space modeling (SSM) in revealing long-range dependencies, it is integrated into the feature pyramid network (FPN) of the fire detection feature extraction network to extract global contextual information. To improve the stationarity of the entire model training, Wise-IoU is used for bounding box regression loss. To evaluate the performance of the proposed MDC-WDA model, we conducted full experiments on two datasets, Fire-Smoke and self made dataset. The experimental results indicated that the MDC-WDA model achieved better results compared to other models, the <i>mAP</i> is 81.2%, which is 1.5% higher than the benchmark model on the Fire-Smoke dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MFDC: Multi-dimensional fusion dynamic convolution based on wind direction aware for fire and smoke detection

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Cuijin Li,
  • Jianyu Liu,
  • Zhong Qu

摘要

Due to variations in wind direction, the shape of fire can undergo significant changes. Traditional convolution, with its fixed weights, cannot adaptively extract image features, making it challenging to handle images with complex backgrounds and unevenly distributed fire of different shapes.Therefore, to address this challenge, we investigated the unique structure of fire and proposed a multi-dimensional dynamic convolutional fire detection method (MDC-WDA) based on wind direction aware. Firstly, a wind direction-aware convolution kernel (WDA) is proposed to adaptively focus on changes in fire and smoke by capturing wind direction, enhancing the model’s perception ability and accurately capturing the features of object structures in images. Secondly, in the feature extraction network CSPDarknet, a multi-dimensional dynamic convolutional network (MDC) is proposed to adaptively change the weights of multiple dimensions of convolution, extract prior knowledge provided by different views, deal with localization difficulties caused by complex scenes and uneven distribution. Finally, utilizing the excellent ability of state space modeling (SSM) in revealing long-range dependencies, it is integrated into the feature pyramid network (FPN) of the fire detection feature extraction network to extract global contextual information. To improve the stationarity of the entire model training, Wise-IoU is used for bounding box regression loss. To evaluate the performance of the proposed MDC-WDA model, we conducted full experiments on two datasets, Fire-Smoke and self made dataset. The experimental results indicated that the MDC-WDA model achieved better results compared to other models, the mAP is 81.2%, which is 1.5% higher than the benchmark model on the Fire-Smoke dataset.