The color, shape, and other appearance characteristics of the flame emitted by different flame engines are different. In order to make a preliminary judgment on the category of the device to which it belongs through s...The color, shape, and other appearance characteristics of the flame emitted by different flame engines are different. In order to make a preliminary judgment on the category of the device to which it belongs through studying exterior characteristics of the flame, this paper uses the flame of matches, lighters, and candles to simulate different types of flames. It is hoped that the flames can be located and classified by detecting the characteristics of flames using the object detection algorithm. First, different types of fire are collected for the dataset of experiments. The mmDetection toolbox is then used to build several different object detection frameworks, in which the dataset can be trained and tested. The object detection model suitable for this kind of problem is obtained through the evaluation index analysis. The model is ResNet50-based faster-region-convolutional neural network(Faster R-CNN), whose mean average-precision(mAP) is 93.6%. Besides, after clipping the detected flames through object detection, a similarity fusion algorithm is used to aggregate and classify the three types of flames. Finally, the color components are analyzed to obtain the red, green, blue(RGB) color histograms of the three flames.展开更多
基金supported by the National Natural Science Foundation of China (61871058)。
文摘The color, shape, and other appearance characteristics of the flame emitted by different flame engines are different. In order to make a preliminary judgment on the category of the device to which it belongs through studying exterior characteristics of the flame, this paper uses the flame of matches, lighters, and candles to simulate different types of flames. It is hoped that the flames can be located and classified by detecting the characteristics of flames using the object detection algorithm. First, different types of fire are collected for the dataset of experiments. The mmDetection toolbox is then used to build several different object detection frameworks, in which the dataset can be trained and tested. The object detection model suitable for this kind of problem is obtained through the evaluation index analysis. The model is ResNet50-based faster-region-convolutional neural network(Faster R-CNN), whose mean average-precision(mAP) is 93.6%. Besides, after clipping the detected flames through object detection, a similarity fusion algorithm is used to aggregate and classify the three types of flames. Finally, the color components are analyzed to obtain the red, green, blue(RGB) color histograms of the three flames.
