PMSGAN: Parallel Multistage GANs for Face Image Translation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Face image translation requires high fidelity in texture details, posing challenges for current methods.
• Artifacts in generated images significantly impact perceived quality.

Purpose of the Study:

• To develop a novel architecture for high-quality face image translation.
• To improve the synthesis of realistic face images by addressing texture detail requirements.

Main Methods:

• Proposed a novel parallel multistage architecture based on Generative Adversarial Networks (PMSGAN).
• Implemented a coarse-to-fine strategy with progressively decreasing spatial resolution inputs.
• Introduced a cross-stage atrous spatial pyramid (CSASP) for inter-stage information fusion.
• Incorporated an attention-based module for refining final activations using multistage decoded outputs.

Main Results:

• PMSGAN progressively learns the translation function through parallel stages.
• The CSASP structure facilitates contextual information exchange between stages.
• The attention module refines generated images using in situ supervised attention.
• Extensive experiments demonstrate PMSGAN outperforms state-of-the-art approaches on face image translation benchmarks.

Conclusions:

• The proposed PMSGAN architecture achieves superior performance in face image translation.
• The novel design effectively handles texture details, yielding high-quality synthetic faces.
• PMSGAN represents a significant advancement in generative adversarial network applications for image synthesis.