TP-GAN UNTUK FRONTALISASI WAJAH VIA ADAPTIVE META-LEARNER-BASED KNOWLEDGE DISTILLATION

Fathin Difa Robbani, Moh. Edi W., S.Kom.,M.Kom., Ph.D dan Arif Nurwidyantoro,M.Cs., Ph.D

2025 | Tesis | MAGISTER KECERDASAN ARTIFISIAL

Face frontalization merupakan tantangan utama dalam pengenalan wajah karena citra dunia nyata sering ditemukan dalam bentuk non-frontal akibat variasi pose, pencahayaan, dan oklusi. TP-GAN (Two-Pathway Generative Adversarial Network) efektif mengatasi masalah ini, namun model tersebut berukuran besar. Penelitian ini mengusulkan pendekatan Meta Knowledge Distillation (MetaKD) pada TP-GAN untuk menghasilkan model student yang lebih ringan tanpa menurunkan kualitas citra. Sebuah meta-learner digunakan untuk menyesuaikan bobot loss secara adaptif pada tiga komponen utama—Pixel Loss, Local Pixel Loss, dan Intermediate Feature Loss—berdasarkan dinamika normalized loss dan gradient norm. Strategi ini memungkinkan pembobotan mandiri (learnable weighting strategy) selama proses distilasi. Hasil eksperimen menunjukkan bahwa student MetaKD mempertahankan kualitas mendekati teacher (SSIM 0,9567, PSNR 38,45 dB) dan melampaui KD konvensional (SSIM 0,9381, PSNR 35,44 dB) dengan kompleksitas komputasi setara (91,83 juta parameter, 1,102 GFLOPs). Analisis pelatihan memperlihatkan bahwa MetaKD membentuk kurikulum adaptif dan stabil melalui mekanisme cross-compensation antar komponen loss. Penelitian ini menunjukkan bahwa integrasi MetaKD pada TPGAN menghasilkan distilasi lebih efektif dengan trade-off optimal antara efisiensi dan kualitas visual, serta berpotensi diterapkan pada arsitektur GAN lain.

Face frontalization remains a major challenge in face recognition, as real-world images are often captured in non-frontal views due to variations in pose, illumination, and occlusion. TP-GAN (Two-Pathway Generative Adversarial Network) has proven effective in addressing this problem; however, the model is computationally heavy. This study proposes a Meta Knowledge Distillation (MetaKD) approach applied to TP-GAN to produce a lighter student model without degrading image quality. A meta-learner is introduced to adaptively adjust the loss weights of three core components—Pixel Loss, Local Pixel Loss, and Intermediate Feature Loss—based on the dynamics of normalized losses and gradient norms. This strategy enables a learnable weighting mechanism during the distillation process. Experimental results show that the MetaKD student preserves image quality close to the teacher (SSIM 0.9567, PSNR 38.45 dB) and outperforms conventional knowledge distillation (SSIM 0.9381, PSNR 35.44 dB) while maintaining comparable computational complexity (91.83 million parameters, 1.102 GFLOPs). Training analysis further indicates that MetaKD forms an adaptive and stable curriculum through cross-compensation among loss components. These findings demonstrate that integrating MetaKD into TP-GAN yields more effective distillation with an optimal trade-off between efficiency and visual quality, and shows strong potential for extension to other GAN architectures.

Kata Kunci : Face Frontalization, TP-GAN, Knowledge Distillation, Meta Learning, Adaptive Loss Weighting, GAN