Geospatial Image Training via Iterative Blurring
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 iteratively blurs and fuzzifies training geospatial images using a mask-based model, generating a probability distribution function and adjusting bias factors to focus on positive features, and applies blurring and fuzzification techniques to guide the model towards learning feature masks, while also modifying the loss function to prioritize accurate identification of positive pixels.
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 preliminary action by pre-processing training images through blurring and fuzzification before feeding them to the model. This transforms sharp binary masks into soft probability distributions, preparing the data in advance to guide the model toward learning positive features rather than degener ating to negative-only learning. The blurring operation is performed beforehand to create a more informative training signal.
Solution Approach 2:
The patent changes parameters by transforming the training data from sharp binary masks to blurred probability distributions. This parameter transformation modifies the input characteristics, converting discrete 0/1 values into continuous probability values between 0 and 1, which provides richer gradient information and prevents the model from overgeneralizing.
2Measurement precision
If the model focuses on learning sparse positive features, then feature identification accuracy improves, but the training process becomes more complex and time-consuming
Solution Approach 1:
By performing blurring and fuzzification as preliminary processing steps before model training, the patent prepares optimized training data that accelerates convergence. This pre-processing creates more informative training samples that guide the model more efficiently, reducing the number of training iterations needed despite the added preprocessing complexity.
Solution Approach 2:
The patent replaces the traditional mechanical approach of using sharp binary masks with a probabilistic approach using blurred distributions. This substitution transforms the training mechanism from discrete threshold-based learning to continuous probability-based learning, which provides smoother gradients and more stable training dynamics.
3Adaptability or versatility
If augmentation techniques are applied to increase known examples, then the model can generalize better to rotation, translation, and scale variations, but the augmentation typically only changes relative appearance and does not bias the model toward positive features
Solution Approach 1:
The patent applies local quality by differentiating the treatment of positive and negative regions through selective blurring. Positive regions receive controlled blurring that preserves their identity while adding probability information, whereas negative regions remain sharp or are treated differently. This localized differential treatment creates a biased training signal that guides the model toward positive features while maintaining transformation generalization.
Solution Approach 2:
The patent performs preliminary blurring and fuzzification on positive examples before augmentation is applied, ensuring that the positive feature bias is established in the training data structure itself. This preliminary transformation of the training signal ensures that even after augmentation transformations, the model receives consistent biased guidance toward positive features.
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. The processor may be configured to blur each of the training geospatial images. The processor may also be configured to iteratively operate a training model to identify a given feature from each of the blurred training geospatial images so that the blurring is reduced with each iteration.