衔接注意力机制与残差ASPP的W-Net工业烟尘图像分割

Connection attention mechanism and residual ASPP of W-Net industrial smoke image segmentation

中文关键词  工业烟尘  图像分割  注意力机制  空洞空间金字塔池化  W-Net

英文关键词  industrial smoke  image segmentation  attention mechanism  atrous spatial pyramid pooling  W-Net

中文摘要

      工业烟尘图像分割是基于烟尘图像监测污染等级判定的重要环节. 针对工业烟尘分割时存在的小目标烟 尘漏检、大目标烟尘误检以及分割结果精度低等问题, 提出了一种结合衔接注意力机制和残差空洞空间金字塔池 化(ASPP)的W-Net网络. 使用衔接注意力机制将两个U-Net网络组合成W-Net, W-Net 能充分利用烟尘的轮廓、位置 信息进行烟尘粗分割和细分割操作, 两次分割能达到更精细的分割效果; 此外, 针对W-Net中的普通卷积功能过于 简单而不能更好地提取烟尘特征的问题, 提出一种兼具残差块和ASPP功能的残差ASPP结构, 同时还能根据大小目 标烟尘的特点进行针对性分割, 分割结果更全面完整. 实验结果表明, 结合衔接注意力机制与残差ASPP的W-Net以 较小的分割效率损失为代价获得了较高的分割精度, Recall, IoU与F-score指标提高了4%～5%, 解决了大小目标烟 尘存在的分割问题, 烟尘的分割效果也优于其他语义分割网络.

英文摘要

      Industrial smoke image segmentation is an important part of pollution level judgment based on the smoke image monitoring. Aiming at the problems of small target smoke missing detection, large target smoke false detection and low precision of segmentation results in industrial smoke segmentation, a W-Net network combining connection attention mechanism and residual atrous spatial pyramid pooling (ASPP) is proposed. Two U-Net networks are combined into a W-Net by using a connection attention mechanism. The W-Net can make full use of the contour and position information of smoke for coarse and fine segmentation, and two segmentations can achieve finer segmentation effect. In addition, aiming at the problem that the common convolution function in W-Net is too simple to extract the characteristics of smoke, an improved ASPP structure with both residual block and ASPP function is proposed, which can also perform targeted segmentation according to the characteristics of large and small target smoke, and the segmentation result is more comprehensive and complete. Experimental results show that the W-Net obtains higher segmentation accuracy at the cost of less segmentation efficiency loss in combination with connection attention mechanism and residual ASPP. The indexes of the Recall, IOU and F-score are improved by 4%～5%, which solves the segmentation problem of large and small target smoke, and the segmentation effect of smoke is also better than other semantic segmentation networks.