The invention discloses a
flame detection method based on an improved RetinaNet network, and the method comprises the steps: S1, collecting N pictures with
flame pictures as a training
data set, and marking flames in the training
data set; S2, a SandGlass module being used for replacing the residual error module, so that an improved RetinaNet network is obtained, and the improved RetinaNet network is recorded as SG-ResNet 50, wherein the SandGlass module comprises a first depth separable
convolution, a first
convolution, a second
convolution and a second depth separable convolution which are connected in sequence; S3, constructing a feature
pyramid network, and adding a segmentation
branch behind each layer of features output by the feature
pyramid network; S4, training the constructed improved RetinaNet network to obtain a trained
flame detection model; and S5, carrying out
flame detection on the obtained video by adopting the
flame detection model obtained in the step S4. According to the invention, the SandGlass module is used for replacing a residual error module of the existing RetinaNet network, so that the
flame detection speed is improved; segmented supervision signals are provided by utilizing the color characteristics of flame, and the flame detection precision is improved.