The paper introduces AM-SegNet, a novel lightweight neural network designed for semantic segmentation and feature quantification of high-resolution X-ray images in additive manufacturing (AM) processes. The authors develop AM-SegNet to address the challenges of processing large volumes of X-ray imaging data generated during AM, such as laser powder bed fusion (LPBF) and directed energy deposition (DED). They establish a large-scale benchmark database containing over 10,000 pixel-labelled X-ray images from various synchrotron experiments, covering different beamlines, materials, and process parameters. AM-SegNet incorporates a lightweight convolution block and a customized attention mechanism, achieving high accuracy (~96%) and processing speed (<4 ms per frame). The model's performance is compared with other state-of-the-art segmentation models, demonstrating superior accuracy and speed. The segmentation results are used for feature quantification and correlation analysis, providing insights into critical features like keyholes and pores. The study also explores the extended applications of AM-SegNet in other advanced manufacturing processes, such as DED and high-pressure die casting (HPDC), validating its versatility and reliability. The proposed method aims to accelerate data processing and analysis in AM and other manufacturing domains, enabling researchers and engineers to gain deeper insights into the underlying physical phenomena. The benchmark database and source codes are made available to support further research and development in this field.The paper introduces AM-SegNet, a novel lightweight neural network designed for semantic segmentation and feature quantification of high-resolution X-ray images in additive manufacturing (AM) processes. The authors develop AM-SegNet to address the challenges of processing large volumes of X-ray imaging data generated during AM, such as laser powder bed fusion (LPBF) and directed energy deposition (DED). They establish a large-scale benchmark database containing over 10,000 pixel-labelled X-ray images from various synchrotron experiments, covering different beamlines, materials, and process parameters. AM-SegNet incorporates a lightweight convolution block and a customized attention mechanism, achieving high accuracy (~96%) and processing speed (<4 ms per frame). The model's performance is compared with other state-of-the-art segmentation models, demonstrating superior accuracy and speed. The segmentation results are used for feature quantification and correlation analysis, providing insights into critical features like keyholes and pores. The study also explores the extended applications of AM-SegNet in other advanced manufacturing processes, such as DED and high-pressure die casting (HPDC), validating its versatility and reliability. The proposed method aims to accelerate data processing and analysis in AM and other manufacturing domains, enabling researchers and engineers to gain deeper insights into the underlying physical phenomena. The benchmark database and source codes are made available to support further research and development in this field.