Abstract
Aiming at the low accuracy of flame detection in complex environment, a new flame detection method based on improved YOLOv3 and gray thermograph is proposed. For the insufficient thermograph of actual flame data sets, a temperature filling algorithm based on radial basis function is proposed. Based on the open data sets, the actual flame image and thermograph are added to build the initial flame data sets, which improves the model training efficiency and generalization performance. The obtained flame gray scale thermograph and R, G, B image are combined to form a four-channel data and input into the improved YOLOv3 network convolution network layer to obtain the multi-channel characteristics of the data. The experimental results on public and self-built data sets show that the algorithm combines gray thermal image to expand image data and improves the detection accuracy of flame recognition.