Multisemantic Path Neural Network for Deepfake Detection

Figure 3

Cropped images of faces are used as input. After multiscale image detail enhancement preprocessing, is obtained, and YCrCb channel separation is performed. The Y channel image is taken as input. Preliminary feature is extracted and put into the microscopic, mesoscopic, and macroscopic semantic channels. , , and are obtained by feature extraction of the three channels. These are fused into the channel and spatial attention modules, and the weight of the three semantic feature maps is allocated. Results are input to the discriminator for classification.