Smoke and fire detection method based on multi-scale dynamic receptive field convolution module enhancement

Through the multi-scale dynamic receptive field convolution module enhanced pyrotechnic detection method, the multi-scale parallel feature extraction architecture and feature splicing are used to solve the problem of insufficient detection of the fire detection model in low-light environments, and efficient and accurate flame and smoke detection are achieved.

CN120339642APending Publication Date: 2025-07-18HUAIYIN INSTITUTE OF TECHNOLOGY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202510427575.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-07
Publication Date
2025-07-18

AI Technical Summary

Technical Problem

The existing fire detection models are difficult to effectively capture flame edges and smoke diffusion characteristics of different sizes in low-light environments, and the calculation complexity is high, which cannot meet the real-time detection requirements.

Method used

The pyrotechnic detection method enhanced by the multi-scale dynamic receptive field convolution module is adopted. Through the parallel branch structure of three heterogeneous convolution kernels, 9×9, 13×13, and 17×17, combined with the initial feature projection and feature splicing, the receptive field is dynamically adjusted to capture the multi-scale visual mode.

Benefits of technology

It improves the accuracy and computing efficiency of fire detection, especially in low-light environments, and is suitable for lightweight model deployment, and can simultaneously capture detailed information and large-scale context information to mitigate the risk of gradient explosion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120339642A_ABST
    Figure CN120339642A_ABST
Patent Text Reader

Abstract

A smoke and fire detection method based on multi-scale dynamic receptive field convolution module enhancement comprises the following steps: 1, performing initial feature projection on an input image, and mapping input features to a low-dimensional feature space through parameterized convolution operation; 2, a multi-scale parallel feature extraction architecture is adopted, and detail information and large-range context information are captured at the same time; focusing high-frequency details of flame edges and textures by small-scale convolution; mesoscale morphological characteristics in the smoke diffusion process are captured through mesoscale convolution; global context information is perceived through large-scale convolution; 3, fusing the features extracted by the multi-scale convolution along the channel dimension through feature splicing; and 4, carrying out feature fusion and channel dimension reduction by adopting 1 * 1 convolution. The multi-scale parallel feature extraction architecture can improve the flame and smoke detection precision under the condition that the cost is not changed or reduced, and is suitable for electric vehicle charging stations, intelligent monitoring systems and other scenes needing high-precision and low-delay fire detection.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Cited By

  • Feature fusion method based on special-shaped convolution kernel

    CN120563989A