DeepErase: A Generative AI Approach for Automated Logo Detection and Removal in Video Content

We present DeepErase, a novel end-to-end framework for automatic logo removal in videos that combines advanced object detection with state-of-the-art inpainting techniques. Our approach leverages a hybrid neural architecture integrating YOLO-based logo detection with Large Mask (LaMa) inpainting, enhanced by a proprietary layer masking system. The framework employs a cascaded pipeline beginning with precise logo detection using a confidence threshold of 0.6, validated through extensive experimentation on the OpenLogo dataset. Multi-layer masking is used to prevent the information loss on nontargeted region of the inpainting through a combination of magenta and white pixel masking layers. Our method significantly reduces computational overhead through efficient frame buffer management and asynchronous processing, making it suitable for fast execution and video creation. The proposed framework not only advances the state-of-the-art in video logo removal but also provides insights into effective combining detection and inpainting architectures for complex video processing tasks.