Abstract: To realize the real-time monitoring of the casing atomization effect of nozzles during the tobacco strip casing process, a real-time visual monitoring system based on deep learning was designed. The system integrates image acquisition module, image processing module and online alarm module to acquire real-time images of nozzle atomization. The image processing module extracts atomization features by morphological processing and an improved lightweight semantic segmentation model (LightBiSeUNet). Based on an asymmetric dual-path parallel architecture comprising a detail branch and a semantic branch, the model effectively suppresses complex interferences, such as the bright background of the cylinder by introducing lightweight depthwise separable convolution (LightConv), Squeeze-and-Excitation (SE) channel attention mechanism, and global average pooling-based global context enhancement (GAP-based GCE) module, achieving adaptive focusing on irregular atomization edges. Experimental results showed that: 1) The LightBiSeUNet model achieved a mean intersection over union (MIoU) of 99.57% with a parameter count of about 0.110 M, enabling precise segmentation of atomization images captured by a high-speed camera frame by frame. 2) Comparisons with U-Net, MobileUNet and BiSeNet further verified its advantages in segmentation accuracy and resistance to overfitting. 3) The application in the industrial production line of Liuzhou Cigarette Factory indicated that the closed-loop control based on high-frequency feedback from this vision system enabled the casing cylinder to intervene in transient anomalies in seconds, reducing the coefficient of variation of casing uniformity from 4.80% to 2.24%. This method enables real-time and accurate monitoring of atomization effect and provides a new approach for real-time and lightweight deployment of industrial image segmentation.