Disclosed herein is an improved mechanism for identifying abusive videos by detecting and removing video overlay content. The mechanism can begin by detecting whether an overlay is present within multiple frames of a video. For example, the mechanism can detect the presence of an overlay by determining gradients between multiple frames of a video and can estimate that the overlay that has been applied to one or more frames of the video based on a median of the determined gradients. For the frames of the video, the mechanism can then remove the overlay from the frames of the video. For example, the mechanism can estimate a clean frame of the video in a gradient-space by subtracting the estimated overlay from an original frame of the video. In continuing this example, the mechanism can then reconstruct a clean image corresponding to the estimated clean frame of the video in gradient-space. Upon obtaining the reconstructed clean image in which the overlay has been removed, the mechanism can apply a classifier to the reconstructed clean image to determine whether abusive content may be present in the reconstructed clean image.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.