Semantic Pixel Distances for Image Editing

Josh Myers-Dean, Scott Wehrwein

Abstract: Many image editing techniques make processing decisions based on measures of similarity between pairs of pixels. Traditionally, pixel similarity is measured using a simple L2 distance on RGB or luminance values. In this work, we explore a richer notion of similarity based on feature embeddings learned by convolutional neural networks. We propose to measure pixel similarity by combining distance in a semantically-meaningful feature embedding with traditional color difference. Using semantic features from the penultimate layer of an off-the-shelf semantic segmentation model, we evaluate our distance measure in two image editing applications. A user study shows that incorporating semantic distances into content-aware resizing via seam carving produces improved results. Off-the-shelf semantic features are found to have mixed effectiveness in content-based range masking, suggesting that training better general-purpose pixel embeddings presents a promising future direction for creating semantically-meaningful feature spaces that can be used in a variety of applications.