Research On The Segmentation Method Of Industrial Smoke And Dust Emission Target Based On Deep Learning

Smoke blackness is an important indicator in industrial pollution monitoring.The pollution level of smoke is assessed based on Ringelmann scale in industry.Computer image recognition technology can be used for industrial smoke antomatic monitoring based on Ringelmann scale.The difficulty and key link in this process is how to segment the smoke target area from the background accurately.The characteristic of industrial smoke is that its shape is not fixed and it is very similar to the cloud,which leads to the inaccurate segmentation results of smoke target area in existing methods in complex backgrounds.Fully Convolutional Networks can perform pixel-level segmentation for smoke target in images,however,the accuracy of smoke target segmentation results in complex backgrounds still needs to be improved.This article has conducted the following research on this issue :(1)Fully Convolutional Networks for industrial smoke target segmentation method based on multi-scale convolution.Aiming at the problem that the FCN can not segment smoke target accurately in complex backgrounds,on the basis of the FCN,the multi-scale convolution operation is used to replace the original single-scale convolution to construct a multi-scale Fully Convolutional Networks,which enhances the feature extraction capability of the model by combining receptive fields of multiple scales.The experimental results show that compared with the original FCN,the method of this chapter has improved the integrity of the smoke segmentation,and the segmentation result is more accurate than before.the Io U metric is increased by up to 5.58%,and the F1-score metric is increased by up to 3.6%.(2)Industrial smoke target segmentation method based on FCN-LSTM.Aiming at the problem that segmentation of smoke target is always affected by the non-smoke elements in the backgrounds,an industrial smoke image segmentation method based on FCN-LSTM is adopted.Based on the FCN for image space feature extraction,Long Short-Term Memory network is used to extract the time information of the image sequence,which distinguishes the moving smoke and the background through the dynamic characteristics of the smoke,and enhances the anti-interference ability in complex bakgrounds.The experimental results show that the improved model has significantly improved the anti-interference ability in complex backgrounds compared to FCN,the model can effectively overcome the interference from the cloud,and solve the problem of interference points in the segmentation results of FCN.Compared with the FCN,the Io U metric is increased by up to 8.04%,and the F1-score metric is increased by up to 5.12%.(3)Industrial smoke target segmentation method based on dynamic weight loss function.Aiming at the problem that the cross entropy loss function cannot sufficiently train difficult and inaccurate classification targets,the loss function with dynamic weight is used to replace the original standard cross entropy loss function.By adding the dynamic weighting factor on the cross entropy to enhance the training of inaccurate classification targets and weaken the training of accurate classification targets.By improving the training efficiency,the accuracy of the model in segmenting smoke in complex backgrounds is improved.The experimental results show that the segmentation result of the improved model with the fusion of dynamic weight loss function is more accurate than the FCN model using the cross entropy loss function,The Io U metric is increased by up to 7.31%,and the F1-score metric is also increased by up to 4.75%.