Abstract: Liquid holdup is one of the most important parameters of gas-liquid two-phase flow. Affected by the image noise, the traditional liquid holdup measurement method based on image processing is difficult to accurately extract the gas-liquid interface, resulting in the great error of the liquid holdup measurement. Hence, a network model based on Deeplab V3+ was established with the semantic segmentation algorithm of deep learning, which was also trained by the image data set obtained from the slug flow pattern acquisition experiment, and then the interface of the gas and liquid flow area in the slug flow was identified and extracted to realize the measurement of the liquid holdup of the gas-liquid two-phase slug flow. The results show that the semantic segmentation model can extract the images of the gas-liquid two-phase flow online, and accurately divide the gas zone, liquid zone and background zone. In addition, the liquid holdup obtained by the top interface extraction is relatively high, while that obtained by the bottom interface extraction is relatively low, but the liquid holdup obtained by the average liquid film thickness is the closest to the true value measured by WMS (Wire-mesh Sensors) grid imaging sensors, with the maximum error less than 10%. The research results provide a new way for real-time monitoring of liquid holdup.