![]() The pupils have near-circular shapes for healthy adults.Ĭomparing with the real faces, we observe that visible artifacts and inconsistencies can be observed in the eye regions of the GAN-generated faces (e.g., StyleGAN2 ). The center of an eye is the iris and pupil, and the white area is the sclera. Concretely, we start with the main anatomic parts of a human eye (see Figure 1(top)). To eliminate these limitations and explore a more robust model, in this work, we propose a new physiological-based method based on pupil shapes. However, when the portrait setting is not obeyed, the method will raise many false positives. However, the proposed method under strict portrait settings such as the light sources or reflectors in the environment are visible to both eyes, and the eyes are distant from the light or reflection source. The work proposes to use the inconsistency of the corneal specular highlights between the two synthesized eyes to distinguish the real and the GAN-generated face images. On the other hand, physical-based methods are proposed to overcome the above limitations by exposing the inadequacy of the GAN synthesis models in representing the human faces interaction with the physical world. However, these methods typically suffer from two significant challenges: the lack of interpretability of the detection results and low robustness of generalization across different synthesis methods due to the over-fitting problem. A recent development on the GAN-faces detection approaches using the Deep Learning model has shown the promising feasibility.
0 Comments
Leave a Reply. |