Computer Vision News - May 2019

Watermark Removal: Watermarks are in widespread use for protecting copyrighted images and video footage. Double-DIP can remove watermarks by treating them as a special case of image reflection, where layers y1 and y2 are the cleaned-up image and the watermark, respectively. In this case, unlike for segmentation, the mask is not set, and inherent transparent layer ambiguity is dealt with by using one of two practical solutions: 1) If only a single watermark is involved -- the user provides a cue by marking the area with the watermark using a bounding box; 2) When there are a few images with the same watermark (2-3 images are usually enough), the ambiguity resolves itself in the training process. The decomposition of a few images sharing a watermark is illustrated below: Conclusion: Double-DIP provides a unified framework for unsupervised layer decomposition of a single image. The proposed framework is applicable across a wide variety of tasks. The framework doesn’t require additional data beyond an input image / video. Despite the methods generality, on some tasks (such as dehazing) it achieves results equal or even better than current state of the art specialized methods in the field. The authors believe that augmenting Double-DIP with semantic/perceptual cues may lead to improved performance on semantic segmentation and other high-level tasks in computer vision. They intend to pursue this in future research. Research 10 Research Computer Vision News …augmenting Double-DIP with semantic/perceptual cues may lead to improved performance on semantic segmentation and other high-level tasks in computer vision.