FCN Uncertainty Quantification via D-S Evidence Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning methods for semantic segmentation cannot effectively quantify model uncertainty, leading to challenges in judging the credibility of neural network decision-making.
Innovation Solution
A semantic segmentation network model uncertainty quantification method based on evidence inference is proposed, which involves constructing an FCN network model, transplanting the D-S theory of evidence, and calculating uncertainty values using the D-S theory of evidence index.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing deep learning methods are used for semantic segmentation, then segmentation results can be obtained, but model uncertainty cannot be quantified and credibility of decision-making cannot be judged
Solution Approach 1:
The patent introduces D-S evidence theory as an intermediary framework between the neural network output and the final decision-making process. The basic probability assignment function acts as a mediator that transforms network outputs into evidential representations, enabling uncertainty quantification without modifying the original segmentation architecture. This intermediary layer preserves the original segmentation functionality while adding credibility assessment capabilities.
Solution Approach 2:
The patent transforms the traditional single-point probability output of neural networks into a distributed evidential representation using basic probability assignment functions. By changing the parameter representation from deterministic probabilities to evidential masses over focal elements, the system enables uncertainty measurement while maintaining segmentation performance.
2Measurement precision
If traditional probability theory is used, then single-point probability assignment is obtained, but it cannot represent uncertainty ranges or conflicting evidence
Solution Approach 1:
The patent segments the sample space into focal elements that can represent different levels of certainty and conflicting hypotheses. By dividing the probability space into distinct focal elements (certain categories, uncertain categories, conflicting evidence), the D-S framework enables precise uncertainty measurement while maintaining computational tractability through the basic probability assignment function.
Solution Approach 2:
The patent demonstrates that the D-S evidence theory framework can handle multiple functions simultaneously: classification, uncertainty quantification, and conflict detection. The same basic probability assignment function serves to represent both certain and uncertain knowledge, making the framework universally applicable to diverse semantic segmentation scenarios without requiring separate mechanisms for each function.
3Reliability
If Bayesian neural networks are used for uncertainty quantification, then predictive uncertainty can be obtained, but calculation amount increases significantly
Solution Approach 1:
The patent employs MC dropout as a computationally efficient approximation method that provides uncertainty estimates without requiring full Bayesian inference. By using dropout during testing as a disposable approximation of Bayesian neural networks, the system obtains predictive uncertainty at a fraction of the computational cost, sacrificing some theoretical rigor for practical efficiency.
4Loss of information
If MC dropout is used for uncertainty measurement, then approximate Bayesian inference is achieved, but large-scale model inference is required increasing time cost
Solution Approach 1:
The patent performs uncertainty quantification as a preliminary step integrated into the forward propagation process itself, rather than requiring separate large-scale inference. By computing the basic probability assignment function during the standard forward pass, the system obtains uncertainty information without additional computational burden, effectively performing the uncertainty measurement action in advance alongside the primary segmentation task.
Data Source
AI summary
A semantic segmentation network model uncertainty quantification method based on evidence inference. The method comprises the steps of constructing an FCN network model, and training the FCN network model by using a training data set to obtain a trained FCN network model for semantic segmentation of image data; transplanting a D-S theory of evidence to the trained FCN network model to obtain a reconstructed FCN network model; and inputting to-be-segmented image data into the reconstructed FCN network model, outputting a classification result of a to-be-segmented image by the FCN network model, and calculating a classification result uncertainty value of each pixel point by using the D-S theory of evidence index.