Abstract: To address the issue of real-time monitoring of the atomization effect of the nozzle on the material liquid during the tobacco lamina feeding process, a real-time visual monitoring system based on deep learning is designed. This system integrates an image acquisition module, an image processing module, and an online alarm module to capture real-time images of nozzle atomization. The image processing module utilizes morphological processing and an improved lightweight semantic segmentation model (LightBiSeUNet) to extract atomization features. Based on an asymmetric dual-path parallel architecture (detail branch and semantic branch), this model effectively suppresses complex interferences such as the bright background of the roller by introducing lightweight depthwise separable convolution (LightConvBlock), Squeeze-and-Excitation (SE) channel attention mechanism, and global average pooling (GAP) enhancement module, achieving adaptive focusing on irregular atomization edges. Experimental results show that the LightBiSeUNet model achieves an average mean intersection over union (MIoU) of 99.57% with only about 0.11M parameters, enabling full and precise segmentation of atomization images captured by a high-speed camera frame by frame. Comparisons with models such as U-Net, MobileUNet, and BiSeNet further verify its superiority in segmentation accuracy and resistance to overfitting. The application in the industrial production line of Liuzhou Cigarette Factory demonstrates that the closed-loop control based on high-frequency feedback from this visual system enables the feeder to intervene in real-time for transient anomalies, reducing the coefficient of variation (CV value) of tobacco lamina feeding uniformity from 4.80% to 2.24%. This method not only achieves real-time precise monitoring of atomization effects but also provides new ideas for the real-time and lightweight deployment of industrial image segmentation.