ARiRTN: A Novel Learning-Based Estimation Model for Regressing Illumination.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Computational color constancy aims to restore object colors under varying illumination.
• Current illumination regression methods face accuracy challenges due to label vagueness from unknown light sources, object properties, and sensor variations.

Purpose of the Study:

• To introduce a novel learning-based estimation model, the Aggregate Residual-in-Residual Transformation Network (ARiRTN), for improved computational color constancy.
• To address accuracy limitations in existing illumination regression approaches.

Main Methods:

• Developed an Aggregate Residual-in-Residual Transformation Network (ARiRTN) architecture.
• Combined Inception and Residual networks, embedding residual networks within a residual network.
• Designed the model with a feature-map group and an ARiRTN operator for simultaneous transformations and concatenation.
• Extended network depth and width through multiple homogeneous branches and increased transformation sets.

Main Results:

• The ARiRTN architecture demonstrated that increased complexity enhances accuracy.
• The combined network structure mitigated overfitting, gradient distortion, and vanishing gradient problems.
• Experimental results on four popular datasets showed superior performance compared to state-of-the-art methods.
• The model exhibited robustness in illuminant and camera invariance.

Conclusions:

• The proposed ARiRTN model significantly improves accuracy in computational color constancy.
• The novel architecture effectively handles challenges related to illumination and sensor variations.
• This approach offers a more robust solution for restoring true object colors in diverse imaging conditions.