Deep Learning-Based Background Removal using MODNet with Green Spill Correction

Abstract - Background removal (portrait matting) is an important pre-processing step for photography, augmented reality, and videography. Traditional chroma-key tools require manual trimaps or controlled lighting, limiting flexibility. We propose an automated deep learning solution that uses the MODNet portrait matting model in a Streamlit-based web application. Users upload a subject image (e.g. human on green/any background) and a new background image; the system runs MODNet to compute an alpha matte, applies green-spill correction, and overlays the foreground onto the new background. The UI includes an interactive canvas for adjusting the subject’s position and a download button for the final composite. We compare MODNet’s performance to other segmentation baselines (DeepLabv3, U²-Net as used in Rembg) in terms of mask quality. Evaluation metrics such as Intersection-over-Union (IoU) and pixel accuracy are used to quantify performance. Results indicate that MODNet produces superior edge delineation and photo-realistic composites compared to standard segmentation approaches, providing an effective and user-friendly background removal tool. . Key Words: Background removal, portrait matting, image segmentation, MODNet, DeepLabv3, U²-Net, Streamlit, interactive interface