Resolving Color Ambiguity in Image Colorization through Caption Guidance and Probabilistic Decoding Altyazi Rehberli?gi ve Olasiliksal Kod Çözümü ile Görüntü Renklendirmede Renk Belirsizli?ginin Giderilmesi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636758
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: Cross-Attention, Image Colorization, Mixture Density Network, Text Conditioning
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
This paper considers grayscale image colorization as an inherently ambiguous mapping problem in which a single luminance value may correspond to multiple plausible colors. To resolve this ambiguity, two complementary mechanisms are investigated: semantic text conditioning via automatically generated image captions, and probabilistic output modeling via a Mixture Density Network (MDN). Two scenarios are considered depending on whether text guidance is incorporated through cross-attention alone or jointly with feature-wise linear modulation (FiLM). By comparing four model variants on a COCO 2017 subset across multiple perceptual metrics, the trade-offs between pixel-level fidelity and distributional realism are characterized. Although multimodal color distributions are typically difficult to model in a regression framework, the combination of caption-guided conditioning and the MDN output head turns out to be effective in reducing color ambiguity.