Accurate screen detection in presentation videos using deep learning

Lecture videos are widely used in classroom and conference environments, where digital slides are frequently displayed on a screen, making screen detection essential for extracting slide areas from presentation videos. This study presents a method for identifying the position of slide areas in video frames by utilizing the You Only Look Once (YOLO) object detection framework. A tailored YOLOv7 model is trained using a labeled dataset that includes frames from presentation videos featuring projected slides. The trained model is subsequently evaluated on unfamiliar images to correctly identify projector screens. The dataset includes more than 2,000 labeled frames, which are increased to 5,000 images by using data augmentation methods. The suggested approach is assessed in comparison to other renowned object detection models. Experimental findings show that the customized YOLOv7 model attains superior accuracy and computational efficiency relative to the standard YOLOv7 and RetinaNet. The results indicate that this method provides a dependable solution for detecting projector screens and can be utilized in different real-world situations.