Fire Object Detection Algorithm Based on Improved YOLOv3-tiny

In view of the poor performance of the commonly used network models on the self-made fire dataset and the real-time requirements for the fire detection, we propose an improved fire object detection model based on the YOLOv3-tiny lightweight network. The specific measures are as follows: Replace the first four pooling layers in the backbone network with the
$3\times 3$
convolutional layers with a step size of 2, increasing network nonlinearity while reducing the information loss caused by the pooling layer; On the basis of the original network detection layer, a new detection layer of
$52\times 52$
size is added to improve the detection effect of small object fire; In order to reduce the detection errors caused by the complex background of the dataset, the SE channel attention module is added after the fusion of different detection layers to strengthen the attention to important features. According to the experimental results, the improved YOLOv3-tiny network has a mean average precision rate of 80.8%, which is 5.9% higher than the original network, which fully proves the effectiveness of the model.