Face Image Translation Based On Generative Adversarial Text

Nowadays,more and more information technology has been applied to various fields,which also reflected in the detection of criminal cases.When the public security organ investigates a case,the criminal analog portrait technology is an important investigation method.The public security organ can draw the appearance characteristics of criminal suspects based on the narrative of eyewitnesses,thus greatly improving the probability of detecting crimes.However,the portrait image is often drawn by a hand-drawn sketch image.Therefore,a technique is urgently required to efficiently and accurately convert the hand-drawn sketch face image into a color face image to restore the facial features of the face image.With the success of GAN in generating clear images,Methods based on GAN have received increasing attention from researchers at home and abroad in processing image translation tasks.In order to promote the translation effect of the face image translation method,we propose a face image translation method based on generative adversarial text(T-GAN).This method uses a Word-LSTM model to extract the text features.The text features and the matched images can produce higher compatibility scores.Then combined with the idea of generative adversarial "games",we designed the generative adversarial text network,and added a third type of input to the discriminator,which was composed of the real image with the unmatched text.Through this method,the training effect of the discriminator is enhanced,and the discriminator is forced to judge whether the generated image conforms to the text description,so that the discriminator can better learn the correspondence between the text description and the image content.Because the number of text descriptions is an important factor in limiting the effect of generating images during the training process,we can generate a large amount of additional text descriptions by interpolating between embeddings of training set.To prove the superiority of our method,the method in this paper and contrast method are tested in the CelebA face dataset.The experimental results show that the proposed method is applicable with image translation results of high quality in all kinds of skin color and hair color face image translation tasks,and the translation effect is better that compared with other image translation methods.