Flame detection fused with feature-weighted attention
Aiming at the problems of low detection accuracy,false detection,and missed detection of flame detection methods based on image processing,a flame detection algorithm that fuses object extraction and feature-weighted attention is proposed in this paper,and a flame dataset is constructed for training the flame detection model.Firstly,the flame and flame color similarities are extracted by the chromaticity separation algorithm in YCrCb color space,In the process of extracting the flame,the outline and color information of the flame are preserved at the same time,thereby filtering out the background information irrelevant to the flame,reducing the model training time and speeding up the network convergence.Secondly,the flame-extracted images are used as input,according to the channel characteristics of the flame-extracted images,the Flame Color-weighted Attention mechanism is constructed in the spatial domain,making the network additionally weight objects other than the background to make the model pay more attention to the flame target and improve the network detection accuracy;Finally,the original model's ordinary convolution is replaced by the more economical Ghost Convolution,which reduces the amount of network computation and reduces the size of the model while outputting feature maps of the same size.The experimental results show that the accuracy and average accuracy of the improved algorithm is increased by 1.4 percentage point and 0.9 percentage point respectively compared with the original algorithm,and the number of parameters is reduced by 17.24％compared with the original model,and the improved algorithm is superior to the current mainstream object detection algorithm in terms of detection accuracy and model size.In the actual detection in different scenarios,the missed detection and false detection of the original model are improved,the classification confidence is also been improved,the detection speed meets the real-time requirements,and the improved detection model can be better applied to fire prevention and control tasks.
public safetyobject detectionobject extractionattention mechanismYOLOv5s algorithm
NETL
NSTL
万方数据
维普
