Core Concepts
Automatic colorization of gray images is improved by optimizing color classes and balancing feature distribution.
Abstract
The article discusses the challenges in automatic colorization of grayscale images with varying object colors and sizes.
Proposes a method to address feature imbalance by transforming color values into discrete color classes.
Experimented on different bin sizes for color class transformation, proposing 532 color classes for classification task.
Introduced a class-weighted function based on true class appearance during training to ensure proper saturation of individual objects.
Proposed a novel object-selective color harmonization method empowered by the Segment Anything Model (SAM).
Presented two new color image evaluation metrics, CCAR and TAR, showing superior performance compared to state-of-the-art models.
Stats
特徴の不均衡を解決するために、クラスごとに重み付けされた関数を提案します。
532色のクラスを分類タスク用に提案します。
セグメントアニシングモデル(SAM)によって強化された新しいオブジェクト選択的カラーハーモナイゼーション手法を提案します。
Quotes
"Unveiling the Spectrum: From Monochrome to a Splash of Hues."
"Our proposed model outstrips other models in visualization, CNR and in our proposed CCAR and TAR measurement criteria."