MSCA_UNet3+: UNet3+ with multi-scale convolutional attention mechanism for industrial soot segmentation

Industrial soot is an important source of air pollution, and the Ringelmann blackness method is currently the most widely and effective means to monitor the pollution level of industrial soot, accurate identification of industrial soot images is the key to the effectiveness of this method. Aiming at the characteristics of simultaneous multi-scale soot targets in soot images, this paper proposes an improved UNet3+ algorithm incorporating a lightweight multi-scale attention mechanism, which embeds a lightweight multi-scale attention mechanism in the network decoding path, enhances the integration of local details and global information by the network decoder, and further improves the network's segmentation capability for multi-scale targets in soot images. Finally, the network is subjected to ablation and comparison experiments on industrial soot datasets, and the experimental results show that UNet3+ with multi-scale attention mechanism can effectively capture multi-scale and multi-directional features in the image without significantly increasing the number of parameters, which improves the segmentation accuracy of the network on soot images.