A Smoke Detection Algorithm Based on Improved YOLO v7 Lightweight Model for UAV Optical Sensors

As a crucial technology for improving the autonomous sensing capability of optical sensors on unmanned aerial vehicles (UAVs), object detection has emerged as a research focus in the UAV patrol inspection. To address the problem of high computational complexity in detection models carried by UAV, a...

Full description

Saved in:
Bibliographic Details
Published in:IEEE sensors journal Vol. 24; no. 16; pp. 26136 - 26147
Main Authors: Sun, Hao-Ran, Shi, Bao-Jun, Zhou, Ya-Tong, Chen, Jing-Hui, Hu, Yun-Long
Format: Journal Article
Language:English
Published: New York IEEE 15.08.2024
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Subjects:
Accuracy
Activation function improvement
Algorithms
Autonomous aerial vehicles
Complexity
Computational complexity
Computational modeling
Fire hazards
Inspection
light weighting
Modules
Neural networks
Object recognition
Optical measuring instruments
optical sensor
Optical sensors
Parameterization
Smoke
smoke detection
Training
Unmanned aerial vehicles
YOLO
You Only Look Once (YOLO) v7
ISSN:1530-437X, 1558-1748
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:As a crucial technology for improving the autonomous sensing capability of optical sensors on unmanned aerial vehicles (UAVs), object detection has emerged as a research focus in the UAV patrol inspection. To address the problem of high computational complexity in detection models carried by UAV, a lightweight network model based on You Only Look Once (YOLO) v7 is designed. Based on this model, a smoke detection algorithm is developed. First, a new multiparameter gated activation function (Better-mish) based on the Mish function is constructed. Second, Conv + batch normalization (BN) + SiLU (CBS) and channel attention (CA) modules are improved based on the Better-mish. The Improved-CA is replaced with the Conv in mix-precision convolution (MPConv) to build CA-MPConv. The improved modules are added to the backbone of the model. Third, the backbone is reconstructed in combination with re-parameterization visual geometry group (RepVGG) module. An improved YOLO v7 lightweight model is proposed, and a smoke detection algorithm based on this model is developed. Finally, several experiments are conducted on Smoke and common object in context (COCO) datasets. The experimental results show three key points as follows: 1) compared with Sigmoid linear unit (SiLU), Mish, and other activation functions, Better-mish can alleviate the gradient disappearance of neural networks in the saturation region; 2) compared with YOLO v7, the improved YOLO v7 model reduces computational complexity by 93.24% and only increases model parameters by less than 2%; and 3) the algorithm can maintain effective recognition of smoke images at low computational complexity. The smoke detection algorithm provides a theoretical basis and an effective algorithm for the early warning of fire hazards in the UAV patrol inspection.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1530-437X
1558-1748
DOI:10.1109/JSEN.2024.3422509