Iteratively Biased Loss Function for Sparse Geospatial Feature Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models struggle to effectively 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 learning outcomes.
Innovation Solution
An image processing system that iteratively applies a bias factor to a loss function based on the number of incorrectly identified pixels, using a mask-based training model and techniques like inverse transform sampling and Shannon entropy to create a probability distribution function, and employs blurring and fuzzification to guide the model towards accurate feature identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional deep learning models are trained with sparse positive examples and abundant negative examples, then the model can be trained with available data, but the model overgeneralizes and degenerates into learning only negative features, failing to identify sparse positive features effectively
Solution Approach 1:
The patent applies local quality by introducing a bias factor that differentially weights positive and negative examples in the loss function. Instead of treating all examples uniformly, the system applies localized quality adjustments by assigning higher weights to sparse positive examples through the bias factor, enabling the model to focus on identifying rare features without being overwhelmed by the abundance of negative examples. This resolves the contradiction by making the learning process sensitive to local feature distributions rather than global statistics.
Solution Approach 2:
The patent employs parameter changes by dynamically adjusting the bias factor parameter during training based on the ratio of positive to negative examples. The bias factor is modified iteratively to optimize the balance between learning from sparse positive features and avoiding overgeneralization from negative features. This parameter adjustment mechanism allows the model to adapt to the imbalanced data distribution and prevent degeneration into learning only negative patterns.
2Adaptability or versatility
If the amount of known example images is increased through augmentation to help the model learn, then the model can generalize better to transformations, but the augmentation typically only changes relative appearance and does not bias the model towards sparse features
Solution Approach 1:
The patent introduces asymmetry by applying different treatment to positive and negative examples through the bias factor in the loss function. While traditional augmentation applies symmetric transformations to all examples, this system applies asymmetric weighting that amplifies the signal from sparse positive examples relative to negative examples. This asymmetric approach enables the model to learn both generalization patterns and sparse feature identification simultaneously, resolving the contradiction between adaptability and reliability.
Solution Approach 2:
The system performs preliminary action by pre-computing the bias factor based on the known ratio of positive to negative examples before the main training process. This preliminary calculation of the bias factor prepares the training framework to handle sparse features effectively from the outset, allowing subsequent augmentation and training to focus on learning robust patterns without losing sensitivity to rare features. The preliminary bias establishment ensures that sparse feature identification capability is built into the learning process from the beginning.
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 operate a training model to identify a given feature from each of the training geospatial images, and to iteratively apply a bias factor to a loss function based upon a number of incorrectly identified pixels for the given feature. The bias factor may be reduced with each iteration.