Fast flame recognition algorithm base on segmentation network

To solve the low recognition rate of the network to flame and keep the accuracy, we propose an Instance segmentation model for recognizing and locating flames more accurate This network is improved based on the deep learning model Mask R-CNN, it introduces four key components:(1) After analyzing the effects of space and channel attention, we used an efficient convolution channel attention. (2) By comparing the convolution kernel size, an optimized dilated convolution is added to the network, (3) To eliminate redundancy, reducing the depth of the backbone while guaranteeing the accuracy of the network. (4) Finally, Adding a flame extraction algorithm behind the head. Compared with Mask R-CNN, the model size is reduced by 16.3MB, and the recognition accuracy of flame is improved by 1.7%, The comparison shows that the network can also greatly improve the recognition effect of small flames.