Learning Model Construction for Scale Estimation in Noisy Images

The patent segments the scale estimation problem by creating multiple specialized learning models, each trained with different loss functions tailored to handle specific noise conditions in low-light environments. This segmentation allows the system to address different aspects of noise interference through dedicated models rather than a single general-purpose model.

The patent changes the parameters of the learning models by employing multiple different loss functions (e.g., L1 loss, L2 loss, Huber loss) with varying characteristics. Each loss function parameterizes the optimization objective differently, enabling the models to adapt to different noise patterns and lighting conditions, thereby improving scale estimation accuracy in noisy indoor environments.

The patent introduces dynamics by implementing a selection mechanism that adaptively chooses the most appropriate learning model based on the characteristics of the input image or verification results. This dynamic selection process allows the system to manage multiple models efficiently, activating only the necessary models for each specific estimation task, thereby balancing accuracy improvement with system complexity management.

The patent employs feedback mechanisms where verification images with known true scales are used to evaluate the performance of different learning models. The correlation between estimated scales and true scales is calculated, and this feedback information is used to select or refine the optimal model, creating a closed-loop system that manages complexity through performance-driven model selection.

The patent applies preliminary action by pre-training multiple learning models with different loss functions using training images before actual scale estimation is needed. The models are prepared in advance and can be quickly evaluated using verification images, significantly reducing the time required during operational phases while maintaining high reliability through pre-established model performance characteristics.

The patent uses partial verification action by selecting a limited number of representative verification images with known scales to evaluate and compare learning models, rather than exhaustively testing all possible images. This partial verification approach achieves sufficient reliability for model selection without incurring excessive time costs, balancing thoroughness with efficiency.