Application of a Convolutional Neural Network for Detection of Ignition Sources and Smoke

The article discusses various methods for detecting ignition sources on aerial photographs. An algorithm based on color filtering and biorthogonal wavelet transform and the Tiny-YOLOv3 convolutional neural network were chosen for research. For the study, training and test datasets were developed. According to the results of experiments, Tiny-YOLOv3 exceeded the algorithm based on color filtering and biorthogonal wavelet transform in detection accuracy. For image processing algorithm AP was 16%. For the Tiny-YOLOv3 with input layer size of 416 × 416, the detection accuracy (AP) of fire and smoke was 56.5% and 31.9%, respectively.