.png){kind=icon}
.png){kind=icon}
Return
KDCC: device-specific color constancy via knowledge distillation and in-domain fine-tuning
.png){kind=icon}
PRE
AI
Z
L
Q
L
L
W
DOI：10.1117/1.JEI.35.2.023035.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In digital imaging, deep learning-based color constancy models face two main challenges: large architectures that hinder camera integration and poor adaptability to specific sensors due to mixed-sensor training data. In addition, acquiring enough same-sensor-labeled data is difficult. To overcome these issues, we propose a lightweight and adaptable framework combining knowledge distillation with sensor-specific fine-tuning. A pretrained teacher model transfers knowledge to a compact student network, creating a general, efficient model. For adaptation, the student model is fine-tuned on target camera-captured images, with similar to 40% labeled by ground-truth illuminants and the remaining similar to 60% served as unlabeled samples. Supervised loss is computed from labeled data, whereas pseudolabels generated by the teacher guide learning from unlabeled data. Confidence-based weighting ensures focus on reliable pseudolabels. Experiments on multiple datasets show that the fine-tuned student significantly reduces the mean angular error-e.g., by more than 23.40% on Cube+-and improves other metrics. The student model also reduces the parameter count by more than 96.38% compared with the teacher, making it lightweight and suitable for integration into image signal processors. This approach effectively addresses issues of model size and sensor adaptability, offering a practical solution to enhance color constancy in resource-constrained imaging systems.
Keywords:
computational color constancy
knowledge distillation
illuminant estimation
image signal processor
image enhancement
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers