DISGUISED FACE PREDICTION IN DEEP CONDITIONED GENERATIVE ADVERSARIAL NETWORK

In this study, we propose Disguise Face Prediction to address the problem of disguised remotion from heavily disguised faces to capture disguised-free faces. Almost all computer vision systems that have to deal with human’s face are often seriously corrupted by disguise accessories and extraneous objects when operating in real-time. The DFP aims to accomplish robust and photo-realistic disguised face reconstruction invariant of disguise, pose and expression; generally, it comprises generative adversarial model design conditioned by non-linear 3D Morphable Model. It begins by regressing shape geometry and expression with a blend of texture information, typically utilizing non-linear 3DMM projection, fitting and rendering to synthesize 2D neutral face for the given disguise face, we carry-out this synthesis for every image in the training set. We thus further introduce, a Deep Conditioned Generative Adversarial Network, by feeding the generator with the synthesized face image and the corresponding disguised face image, besides, we design two discriminators which elaborate on local and global face attributes of the generated face to get the final refined result