A novel fully convolutional network based on marker-controlled watershed segmentation algorithm for industrial soot robot target segmentation

Soot blackness is an important index in industrial pollution monitoring. Aiming at how to effectively segment soot target area from the background by using computer image recognition technology in the automatic monitoring of Ringelmann scale, and the industrial soot is characterized by unstable shape and high cloud similarity, which leads to inaccurate results of smoke segmentation with traditional existing methods under complex scenes. This paper proposes a novel fully convolutional network (FCN) based on marker-controlled watershed segmentation algorithm for industrial soot target segmentation. Firstly, the marker-controlled watershed algorithm is used to extract the connected regions in the soot image, and the corresponding target positions of the original image are marked and normalized to extract the candidate soot regions. The marker-controlled watershed method can greatly reduce the detection time of candidate regions and the recognition time of FCN. The extracted candidate regions are input into the improved FCN for feature extraction and prediction. Combined with multi-scale convolution operation, the feature extraction ability of FCN network is enhanced. By adding dynamic weights on the basis of cross entropy, the training of inaccurate regions is enhanced and the accuracy of segmentation is further improved. Experimental results on a real industrial soot emission image data sets show that the proposed method is more accurate than the original models under the complex scene for smoke segmentation. The F1 measurement and IoU indexes are significantly improved than that of state-of-the-art segmentation methods, which exceeds 87% and 85% respectively.