Weakly paired image style transfer method based on pose self-supervised generative adversarial network

A weakly paired image style transfer method based on a pose self-supervised generative adversarial network, the following steps are provided:

• S1:
• obtaining an original data set consisting of image sample pairs, each image sample pair comprising a first source image O1 and a first target image T1 that have both a pose difference and a style difference;
• S2:
• for each image sample pair in the original data set, performing three-degree-of-freedom transformation Pr respectively on the first source image O1 and the first target image T1, generating a second source image O2 and a second target image T2, respectively, and forming a sample combination consisting of four images, wherein the three-degree-of-freedom transformation Pr comprises three randomly generated pose transformations comprising horizontal and vertical translation and rotation, and the three-degree-of-freedom transformation PR adopted by the first source image O1 and the first target image T1 in the same image sample pair remains the same;
• S3:
• inputting the sample combination in the original data set as a training sample into a generative adversarial network, performing style transformation on the first source image O1 and the second source image O2 respectively by the generative adversarial network to generate a third target image FakeT1 and a fourth target image FakeT2, and training the generative adversarial network by minimizing a total loss function of the inputted training sample, so as to obtain a weakly paired image style transfer network,
• wherein in the total loss function, a loss function L of a single training sample consists of three parts:
• L=LD+LS+LM,
• where a loss of a first part is LD=(T1−FakeT1)+(T2−FakeT2), a loss of a second part is LS=P3−P4, and a loss of a third part is LM=Pr−Pe,
• in the formula:
• P3 is a relative pose between the third target image FakeT1 and the first target image T1, P4 is a relative pose between the fourth target image FakeT2 and the second target image T2, and P e is a relative pose between the third target image FakeT1 and the fourth target image FakeT2;
• and
• S4:
• inputting an image to be stylized into the weakly paired image style transfer network to obtain a stylized image having an unchanged pose.