A Smoke and Flame Detection Method Using an Improved YOLOv5 Algorithm

Abstract

The complex background scenes in traditional fireworks detection methods make flame identification challenging and complicated. This paper focuses on improving the detection efficiency and accuracy of flame disasters. First, the data augmentation strategy and label smoothing are used to preprocess the sample set, which solves the over-fitting problem caused by the insufficient number of samples. Second, we add Convolutional Block Attention Module (CBAM) before each backbone classifier, to compress and re-weight the input features from two independent channel and space dimensions. By focusing on smoke and fire's feature information, the ability of desired feature extraction is strengthened. Third, the Focal loss function is utilized to enhance the weights of complex samples. Consequently, the imbalance problem about positive and negative samples in single-stage detection, and the high proportion of easy-to-separate samples in the loss function are both resolved. Experimental examples demonstrate that the proposed network is easy to converge and expand, which guarantees detection accuracy and satisfies detection speed requirements.