Research on Improved YOLOv3 Fire Detection Based on Enlarged Feature Map Resolution and Cluster Analysis

Abstract

Fire detection can effectively prevent the occurrence of fire. For the current fire detection methods, traditional image processing techniques such as grayscale image processing and feature extraction processing have poor anti-interference ability, weak generalization ability, and the detection results are more sensitive to data fluctuations. At present, based on the concept of deep learning, the proposed convolutional neural network processing of extracted image features has been widely used. On this basis, an improved YOLOv3 fire detection algorithm is proposed in this paper: The K-Means++ algorithm is used for clustering analysis to obtain the corresponding anchor boxes dimension, which reduces the error detection rate caused by the bounding boxes not matching the label; secondly, the resolution of the feature image is improved and the receptive field is enlarged; the image data are sharpened and the contrast is enhanced to make the data features more prominent. The experimental results show that the detection precision of the method is 97.7%, the recall rate is 98.5%, and the fps is 19, effectively solving the problem of high error detection rate and high missing rate of traditional image processing and general CNN networks for suspected flame objects.