Multispectral Sensor Color Conversion via Matrix Optimization

The patent replaces the complex MLP neural network system with a simplified linear transformation system using a 3x3 matrix. Instead of using a deep learning model with multiple layers and activations, the invention uses a direct mathematical transformation: T_ref = M_ref × S_ref, where T_ref is the reference color space, M_ref is the 3x3 transformation matrix, and S_ref is the sensor signal. This substitution dramatically reduces computational complexity while maintaining color conversion capability.

The patent changes the approach from learning complex non-linear parameters in an MLP to determining simple linear parameters in a 3x3 matrix. The transformation matrix parameters are derived directly from the sensor's spectral sensitivity characteristics and the target color space definition, rather than being learned from large training datasets. This parameter simplification reduces both computational burden and memory requirements.

The patent incorporates a feedback mechanism through the loss function that includes both a data fidelity term and a regularization term: L = ||T_ref - M_ref × S_ref||² + λ||M_ref||². The regularization term provides feedback that penalizes overly complex transformation matrices, preventing overfitting. This feedback loop ensures that the optimized matrix maintains good generalization capability while still fitting the training data adequately.

The patent applies partial optimization by not seeking the perfect fit to training data, but rather a sufficient fit that generalizes well. The regularization parameter λ controls the balance between fitting accuracy and model simplicity. By accepting a slightly suboptimal fit to training data in exchange for improved generalization, the method avoids overfitting while maintaining practical accuracy for color conversion applications.