Hierarchical CNN Image Segmentation for Uncertain Pixel Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in interpreting imperfect sensor inputs, particularly in image segmentation, which can limit their decision-making capabilities and necessitate unnecessary default actions.
Innovation Solution
A hierarchical image segmentation scheme using a convolutional neural network (CNN) that allows classification of pixels based on a parent level even when lower-level classification is uncertain, enabling safer decision-making by relying on higher-level classifications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image classification techniques are used, then the system can provide definitive classifications, but it requires default actions when classification certainty is insufficient, reducing operational flexibility
Solution Approach 1:
The classification hierarchy is segmented into multiple levels (parent level and child level), where each level handles different granularity of classification. The parent level provides coarse-grained classifications that can be used independently when child level certainty is insufficient, while the child level provides fine-grained classifications when available. This segmentation allows the system to adapt its decision-making granularity based on input quality.
Solution Approach 2:
The system performs partial classification action by utilizing only the parent level classification when child level certainty is insufficient, rather than requiring complete classification at all levels. This allows safe decision-making with partial information (parent level only) when full information (child level) is not available, reducing unnecessary default actions.
2Measurement precision
If fine-grained classification is required for all pixels, then decision precision is improved, but the system resorts to default actions more frequently when sensor inputs are imperfect
Solution Approach 1:
The classification system dynamically adjusts its granularity based on input quality. When sensor inputs are perfect, the system uses fine-grained child level classifications. When sensor inputs are imperfect, the system dynamically falls back to coarser parent level classifications, maintaining operational reliability while adapting precision to actual conditions.
Solution Approach 2:
The system prepares multiple classification levels in advance (parent and child levels) so that when classification certainty is insufficient at the fine-grained level, there is already a coarser classification available to cushion the impact and enable continued operation without default actions.
3Productivity
If higher-level classifications are used when lower-level classification is uncertain, then unnecessary default actions are reduced, but the system complexity increases
Solution Approach 1:
The classification system is structured as nested levels where child level classifications are nested within parent level classifications. Each pixel classification result at the child level is associated with its corresponding parent level classification, creating a nested structure that allows automatic fallback from fine-grained to coarse-grained classifications when needed.
4Measurement precision
If the system requires high certainty for all pixel classifications, then classification accuracy is maintained, but decision-making capability is limited by frequent default actions
Solution Approach 1:
The classification requirement is segmented into optional parent level and child level classifications. The system can achieve decision-making capability using parent level classifications alone when child level certainty is insufficient, while maintaining classification accuracy through probabilistic confidence measures at each level.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
In one aspect, hierarchical image segmentation is applied to an image formed of a plurality of pixels, by classifying the pixels according to a hierarchical classification scheme, in which at least some of those pixels are classified by a parent level classifier in relation to a set of parent classes, each of which is associated with a subset of child classes, and each of those pixels is also classified by at least one child level classifier in relation to one of the subsets of child classes, wherein each of the parent classes corresponds to a category of visible structure, and each of the subset of child classes associated with it corresponds to a different type of visible structure within that category.