Iteratively Biased Training Model for Sparse Geospatial Feature Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models struggle to effectively learn and identify sparse features in geospatial images due to overgeneralization and degeneration into learning only 'negative' features, especially when the number of positive examples is significantly less than negative examples, leading to poor performance in feature classification.
Innovation Solution
The system employs a mask-based training model that iteratively generates a probability distribution function based on a bias factor reduced with each iteration, along with blurring and fuzzification techniques, to bias the model towards learning sparse features by adjusting the relative representation of positive and negative pixels and applying a bias to the loss function, ensuring accurate feature identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional deep learning models are trained with sparse positive examples, then the model can be trained with limited data, but the model overgeneralizes and degenerates into learning only negative features
Solution Approach 1:
The patent applies parameter changes by modifying the loss function with a bias factor that dynamically adjusts during training. The bias factor is initialized based on the ratio of positive to negative examples and is gradually reduced over training iterations, allowing the model to initially focus on sparse positive features and then generalize properly. This resolves the contradiction by changing the training parameter (bias factor) to enable effective learning with sparse data while maintaining classification accuracy.
Solution Approach 2:
The patent implements preliminary action by pre-calculating the bias factor before training based on the distribution of positive and negative examples. This preliminary configuration of the loss function bias ensures that the model is properly oriented to handle sparse positive examples from the start of training, preventing overgeneralization before it occurs.
2Quantity of substance
If data augmentation is used to increase training examples, then the model can learn from more data, but the augmentation only changes relative appearance and does not bias the model towards sparse features
Solution Approach 1:
Instead of relying solely on data augmentation, the patent changes the loss function parameter by introducing a bias factor that directly addresses the imbalance between positive and negative examples. This parameter change in the objective function provides more effective guidance than appearance-based augmentation, improving feature identification precision by explicitly biasing the model towards sparse positive features during training.
3Reliability
If the model is biased towards positive features during training, then sparse features can be learned effectively, but the model may overfit to the bias and fail to generalize
Solution Approach 1:
The patent applies dynamics by making the bias factor a dynamic parameter that changes during training. The bias factor is initialized to compensate for sparse positive examples but is gradually reduced over training iterations according to a scheduled decay. This dynamic adjustment allows the model to initially focus on learning sparse features effectively and then transition to generalization, resolving the contradiction between biased learning and adaptability.
Solution Approach 2:
The patent implements periodic action through the scheduled reduction of the bias factor during training. The bias is applied in a periodic manner where its strength decreases over time, creating phases of biased learning followed by phases of more balanced learning. This periodic modulation enables the model to learn sparse features while maintaining generalization capability.
Data Source
AI summary
An image processing system may include a processor and an associated memory configured to store training data that includes training geospatial images. Each training geospatial image may include pixels. The processor may be configured to operate a training model to identify a given feature from each of the training geospatial images, and to iteratively generate a probability distribution function based upon a number of pixels corresponding to the given feature and also based upon a bias factor being reduced with each iteration.