A Dual-Domain Perceptual Framework for Generating Visual Inconspicuous Counterparts

For a given image, it is a challenging task to generate its corresponding counterpart with visual inconspicuous modification. The complexity of this problem reasons from the high correlativity between the editing operations and vision perception. Essentially, a significant requirement that should be emphasized is how to make the object modifications hard to be found visually in the generative counterparts. In this article, we propose a novel dual-domain perceptual framework to generate visual inconspicuous counterparts, which applies the perceptual bidirectional similarity metric (PBSM) and appearance similarity metric (ASM) to create the dual-domain perception error minimization model. The candidate targets are yielded by the well-known PatchMatch model with the strokes-based interactions and selective object library. By the dual-perceptual evaluation index, all candidate targets are sorted to select out the best result. For demonstration, a series of objective and subjective measurements are used to evaluate the performance of our framework.