关键词: 工业烟尘, 图像分割, W-Net, 注意力机制, 焦点损失

Abstract: Industrial smoke image segmentation is the basis of smoke pollution level monitoring.Aiming at the problems of missing detection,false detection and low segmentation accuracy caused by small size and thin edge of small target smoke and the poor effect of U-Net in extracting small target smoke features,a smoke image segmentation network based on Improved Attention W-Net (IAW-Net) was proposed.By using the attention mechanism,the U-Net was extended to W-Net,and an improved attention mechanism was introduced based on W-Net,which enhanced the characteristics of small target smoke.In addition,the focus loss was improved according to the characteristics of small target smoke,which increased the proportion of small target smoke segmentation.The experimental results showed that the IAW-Net could focus more on the segmentation of small target smoke without affecting the large target smoke segmentation,thus improving the overall segmentation ability of smoke image,and had better segmentation effect compared with the existing semantic segmentation network.

Key words: industrial smoke, image segmentation, W-Net, attention mechanism, focus loss