<p>Color constancy is a crucial visual ability that allows humans to perceive consistent colors under varying illuminations. However, cameras often struggle with this under multi-illuminant scenes, resulting in local color casts in images. This paper addresses the challenge of multi-illuminant color constancy, aiming to estimate an illuminant map from an image. Existing methods rely either on the unverified assumption of smooth distribution, or on additional hand-tuned parameters, and prior knowledge about the illuminant, constraining their practical application. Through frequency analysis, we reveal that the illuminant map can contain high-frequency components, violating the smoothness assumption, and as its scale decreases, its frequency distribution shifts toward low-frequency regions. Inspired by these findings, we propose a network comprising both a low-frequency branch and a high-frequency branch. The two branches use U-Nets to estimate low- and high-frequency illuminant maps from small- and large-scale images, respectively. To preserve high-frequency features within the high-frequency branch, we embed a high-frequency preservation gating module (HFPGM) within the skip connections to filter and retain high-frequency features. Furthermore, we embed a cross-scale high-frequency preservation gating module (CSHFPGM) between the two branches. This module leverages the feature discrepancy between the decoding paths of both branches as a constraint and filters the features output by the high-frequency branch decoder. Our method achieves mean angular errors of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1.96^{\circ }\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1</mn> <mo>.</mo> <msup> <mn>96</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(1.80^{\circ }\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1</mn> <mo>.</mo> <msup> <mn>80</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation>, and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(1.68^{\circ }\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1</mn> <mo>.</mo> <msup> <mn>68</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation> on the three subsets of LSMI and <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(1.33^{\circ }\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1</mn> <mo>.</mo> <msup> <mn>33</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation> on the CUBE++ dataset. Our codes are publicly available at: <a href="https://github.com/AirF22/MSN">https://github.com/AirF22/MSN</a>.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Enhancing multi-illuminant color constancy through multi-scale estimation and high-frequency preservation

  • Hang Luo,
  • Rongwei Li,
  • Jinxing Liang

摘要

Color constancy is a crucial visual ability that allows humans to perceive consistent colors under varying illuminations. However, cameras often struggle with this under multi-illuminant scenes, resulting in local color casts in images. This paper addresses the challenge of multi-illuminant color constancy, aiming to estimate an illuminant map from an image. Existing methods rely either on the unverified assumption of smooth distribution, or on additional hand-tuned parameters, and prior knowledge about the illuminant, constraining their practical application. Through frequency analysis, we reveal that the illuminant map can contain high-frequency components, violating the smoothness assumption, and as its scale decreases, its frequency distribution shifts toward low-frequency regions. Inspired by these findings, we propose a network comprising both a low-frequency branch and a high-frequency branch. The two branches use U-Nets to estimate low- and high-frequency illuminant maps from small- and large-scale images, respectively. To preserve high-frequency features within the high-frequency branch, we embed a high-frequency preservation gating module (HFPGM) within the skip connections to filter and retain high-frequency features. Furthermore, we embed a cross-scale high-frequency preservation gating module (CSHFPGM) between the two branches. This module leverages the feature discrepancy between the decoding paths of both branches as a constraint and filters the features output by the high-frequency branch decoder. Our method achieves mean angular errors of \(1.96^{\circ }\) 1 . 96 , \(1.80^{\circ }\) 1 . 80 , and \(1.68^{\circ }\) 1 . 68 on the three subsets of LSMI and \(1.33^{\circ }\) 1 . 33 on the CUBE++ dataset. Our codes are publicly available at: https://github.com/AirF22/MSN.