Image Sensor Parameter Optimization Using Score-Based Evaluation

The patent divides the parameter optimization process into two distinct phases: a learning phase where the neural network learns the relationship between sensing parameters and image quality scores from training data, and an application phase where the learned model directly evaluates and optimizes parameters. This segmentation allows the system to handle parameters where gradient calculation is not possible by relying on the pre-trained evaluation model rather than real-time gradient computation.

The patent introduces an intermediate score evaluation model (trained neural network) that acts as a mediator between the sensing parameters and the final image quality assessment. Instead of directly optimizing parameters through gradient-based methods, the system uses this intermediate model to evaluate potential parameter settings and guide the optimization process, enabling precise optimization even when direct gradient calculation is infeasible.

The patent creates a universal optimization framework that can handle multiple types of sensing parameters (both differentiable and non-differentiable) through a single system architecture. The score evaluation model serves multiple functions: it evaluates image quality for parameter optimization, guides the search for optimal parameters, and adapts to different parameter types without requiring different optimization methods, thus achieving both precision and versatility.

The patent transforms the optimization approach by changing from direct gradient-based parameter adjustment to a score-guided parameter search. By using the trained neural network to evaluate scores for different parameter settings and then optimizing based on these scores (through methods like grid search, random search, or other score-based optimization), the system can effectively optimize any parameter type regardless of whether gradients are available.