基于RetinaNet深度学习模型的火焰检测研究

Flame Detection Based on RetinaNet Deep Learning Model

鉴于卷积神经网络能够自动学习和获取图像特征,提出一种基于RetinaNet模型的火焰检测算法.首先RetinaNet在残差网络之上使用特征金字塔网络作为骨干网计算和生成丰富的卷积特征金字塔,然后通过分类子网络和边框预测子网络分别对骨干网的输出进行分类和回归,通过调整训练策略和参数,最后在自建数据集上使用该算法得到的火焰检测模型实现了实时的端到端火焰识别与定位,对复杂目标背景下的小火点检测也保持较高的检测准确率,对火灾初期的检测预警有一定的实用意义.

Because the convolutional neural network can automatically learn and acquire image features, in the report, a flame detection algorithm based on RetinaNet model was proposed. Firstly, on the top of a feed forward ResNet architecture, RetinaNet uses the feature pyramid network as the backbone network to generate a rich convolution feature pyramid;Secondly, by adjusting the training strategy and parameters, the classification subnet and box subnet was used to classify and regress the output of the backbone network separately;Lastly, the flame detection model based on RentinaNet was used to realize the real-time and end-to-end flame detection on the built data set, which maintains a high detection accuracy rate for small flame detection under complex target background, and has certain practical significance for early detection and early warning of fire.