Industrial Smoke Target Segmentation Based on Fully Convolutional Networks with Multiscale Convolution and Dynamic Weight Loss Function

Abstract

Smoke blackness is an important indicator in industrial pollution monitoring. Aiming at how to use computer image recognition technology to effectively segment the target area of smoke from the background in the automatic monitoring of Ringelmann scale, and the characteristics of industrial smoke and dust with unfixed shapes and high cloud similarity, leading to the inaccurate results of existing methods to segment smoke in complex scenes. An improved FCN model was proposed, combined multi-scale convolution operations to enhance the feature extraction ability of the network. Dynamic weights were added to the basis of cross entropy to enhance the training of inaccurate points, and the accuracy of segmentation was further improved. The experimental results with the actual factory smoke emission image data set shows that the proposed model is more accurate than the other models in the complex scene, the F 1 -score and the IoU metrics are significantly improved.

Video segmentation of industrial smoke based on dynamic fully convolutional network-Gaussian mixture model and multi-scale fusion attention module

“…Image segmentation methods can be categorized into traditional image processing techniques and deep learning methods 6 . In the field of industrial smoke segmentation, traditional digital image processing techniques include texture analysis, 7 , 8 threshold segmentation, 9 frame difference, 10 and mixed Gaussian background modeling 11 . However, these methods can only extract low-level features such as color, spatial structure, and texture for image segmentation, and they often involve long processing times.…”