Transform-based lightweight early fire detection method

Abstract

The invention discloses a Transform-based lightweight early fire detection method. The method comprises the steps of establishing an indoor and outdoor multi-scene fire initial image sample data set; the method comprises the following steps: designing a Transform-based lightweight backbone network, introducing a linear attention enhancement mechanism into a Transform structure, introducing locality into a feed-forward network through sequence and image conversion and depth separable convolution to realize global and local feature processing of an image, and performing down-sampling through an inverse residual block to obtain feature maps with different resolutions; and further enhancing feature extraction through feature enhancement and a multi-scale feature fusion structure, and finally performing detection in the mixed feature map to obtain a flame target detection result. According to the method, the advantages of Transform and the convolutional neural network are combined, while feature extraction is optimized, network parameters and calculation amount are reduced, a lightweight detection model is constructed, high detection precision is ensured, high detection speed is realized, and multi-scene fire early-stage target detection can be well realized.

Description

technical field [0001] The invention belongs to the field of computer vision, in particular to a lightweight early fire detection method based on Transformer. Background technique [0002] Fire is a kind of uncontrolled burning, which causes great harm to public safety, social development and natural ecology. Fire prevention and timely response are the most effective ways to minimize fire hazards. Once the fire spreads, it will cause huge harm to life safety and social property. Therefore, setting up monitoring equipment in various places to monitor the fire situation as soon as possible, notify in time, and fight the fire in the early stage is the most effective way to deal with the fire. Greatly reduce the loss caused by fire. [0003] Traditional fire monitoring equipment is mostly based on temperature or smoke particle sensors. This sensor detection method takes a long time and is easily affected by factors such as space area, height, dust, airflow speed, etc., and can...

AI Technical Summary

Problems solved by technology

In some direct detection methods, the fire targets in the data set are large and the flame color is obvious. When the model is used to detect early fires, the detection effect is poor due to the different detection ranges.

Although the convolutional neural network can realize fire detection, it still has certain limitations in modeling global context information and early fire detection accuracy. The Transformer structure can effectively solve this problem

In recent years, the research on applying Transformer to the field of vision has been emerging, but Transformer has a large amount of calculation and slow detection speed, and the detection effect on small targets is not good.

Images