On the Optimization of Regression-Based Spectral Reconstruction.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Spectral reconstruction (SR) aims to recover hyperspectral data from RGB images.
• Current SR algorithms, especially Deep Neural Networks (DNNs), are evaluated using Mean Relative Absolute Error (MRAE).
• Traditional regression methods for SR are often trained with Root Mean Square Error (RMSE) and use spectrum-level regularization, which is suboptimal for MRAE evaluation.

Purpose of the Study:

• To reformulate regression-based SR algorithms to directly optimize relative error metrics during training.
• To implement a per-channel regularization strategy tailored to MRAE evaluation.
• To improve the performance of regression-based SR methods and reduce the performance gap with DNNs.

Main Methods:

• Formulated regression models to minimize both ℓ2 and ℓ1 relative error metrics (MRAE) during training.
• Introduced a per-channel regularization approach for solving ill-posed linear systems in SR.
• Compared the performance of modified regression methods against established SR algorithms and DNNs using MRAE.

Main Results:

• Modified regression algorithms achieved up to a 14% improvement in mean performance and up to 17% in worst-case performance (MRAE).
• The best-performing regression model significantly narrowed the accuracy gap with leading DNN models to approximately 8% mean accuracy.
• Per-channel regularization and relative error optimization proved effective for regression-based SR.

Conclusions:

• Optimizing regression models for relative error metrics and employing per-channel regularization are effective strategies for enhancing spectral reconstruction.
• These advancements make simpler regression methods more competitive with complex DNNs for SR tasks.
• The study provides a more efficient and accurate approach to spectral reconstruction using modified regression techniques.