Image editing with varying intensities of processing

摘要

Current image editing algorithms prevailingly involve specific processing intensities. After the network training phase, the definite mapping is carried on testing images, which leads to under- or over-editing in certain cases. However, for many real problems, having access to diverse intensities of processing of the output is preferable. The hypothesis is that the part of parameters that correspond to the intensity of processing of the network are implicitly embedded in the whole parameter set and can be explicitly extracted and utilized, even with limited supervision. In this paper, we introduce the concept of Intensity of Processing, termed as IP, and propose image editing with Varying Intensities of Processing, termed as VIP. Given the input and corresponding fully-edited output, we would like to not only figure out the particular mapping from input to output, but also be able to obtain multiple continuous intermediate states with in-between intensities of processing. Refer to the network that deals with the original image editing operation as Base Network. By adding an extra module, which we call Condition Encoder, to condition and regulate the Base Network, we make the whole image editing process controllable. Moreover, we develop a triplet loss and construct the sequentiality of generated intermediate states with limited supervision. A significant advantage of our method is that it does not require groundtruth images of the middle states during training, which substantially relaxes the range of applicable problems. Extensive experiments demonstrate the effectiveness and potentiality of the proposed framework; discussion and future work are also provided.