ClsGAN: Selective Attribute Editing Model based on Classification Adversarial Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Recent advancements in attribution editing leverage encoder-decoder structures and Generative Adversarial Networks (GANs).
• Generating high-quality images with precise attribute transformation remains a significant challenge.
• Existing methods struggle with maintaining photo-realism while accurately transferring attributes.

Purpose of the Study:

• To propose a novel selective attribute editing model, ClsGAN, that achieves a balance between attribute transfer accuracy and photo-realistic image generation.
• To address the issue of original attribute influence in edited images caused by skip-connections in encoder-decoder architectures.
• To enhance the accuracy of attribute transfer in generated images.

Main Methods:

• Developed a classification adversarial network (ClsGAN) for selective attribute editing.
• Introduced an upper convolution residual network (Tr-resnet) for selective information extraction from source images and target labels.
• Incorporated an attribute adversarial classifier (Atta-cls) to guide the generator by identifying attribute transfer defects.

Main Results:

• ClsGAN demonstrates superior performance compared to state-of-the-art approaches on the CelebA dataset.
• The model achieves a favorable balance between image quality and attribute transfer accuracy.
• Ablation studies confirm the effectiveness of the proposed Tr-resnet and Atta-cls components.

Conclusions:

• The proposed ClsGAN model effectively addresses the challenges in high-quality attribution editing.
• The integration of Tr-resnet and Atta-cls significantly improves both image realism and attribute transfer precision.
• ClsGAN represents a promising advancement in the field of image attribute editing.