A fire detection method based on deep learning

Abstract

A fire detection method based on deep learning belongs to the field of fire detection in image processing technology, and includes the following steps: receiving an input image; and pooling operation to extract fire features from the input image, and use skip connections to fuse low-level features and high-level features; use sliding windows of different sizes to extract feature vectors from the fire feature map to predict the fire candidate frame; Region deformable pooling is performed on each fire candidate frame; deep features are input into the fully connected layer for classification, prediction and localization of fires. The invention fuses shallow features with high-level features, and utilizes a convolutional neural network with geometric deformation capability, thereby effectively solving the false detection rate of small flame detection for scenes similar to fire in the background and in the early stage of fire spread. high and low accuracy.

Description

technical field [0001] The invention belongs to the field of fire detection in image processing technology, and in particular relates to a fire detection method based on deep learning. Background technique [0002] The rapid development of deep learning in recent years has enabled more and more artificial intelligence applications to be realized. With convolutional neural networks as the core, convolutional neural networks (CNNs) have been widely used in computer vision tasks, such as object classification. , target segmentation, and target detection. Therefore, many scholars and researchers have also applied deep learning to the field of fire detection recently, including traditional convolutional neural network structures such as AlexNet, VGGNet, GoogleNet and ResNet for fire image classification, as well as R-CNN, Fast R-CNN, Faster R-CNN, YOLO, SSD and other models are used for location detection of fire images. [0003] Although the above deep learning models have imp...

AI Technical Summary

Problems solved by technology

[0003] Although the above deep learning models have improved the accuracy of flame detection on a certain basis, because deep learning requires a large amount of data sets, the data sets used are usually large fire pictures searched on Google and Baidu, so for some specific scenarios Under the background, if the background is similar to the flame, and the detection accuracy of the small flame before the fire spread becomes larger, etc., it is still very low. Therefore, in view of these problems, it is proposed to use the skip connection method of low-level features and high-level features for feature fusion, and according to The characteristic of flame shape changing during the fire occurrence uses deformable convolution for feature extraction, which effectively improves the accuracy of flame detection and greatly reduces the misrecognition rate, and is more effective for flames and small flames in different scenes. good detection effect

Images