GPR Image Classification via Segmented Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Ground-penetrating radar (GPR) images are challenging to classify accurately due to their complexity and diversity, making it difficult to extract features effectively, which hampers the performance of existing image classification models.
Innovation Solution
The method involves separating the training processes of feature value extraction and image classification models, using K-means clustering to determine a representative feature value with a minimum distance from the centroid feature value for each label, and initializing the feature value extraction model with a label value before learning the representative feature value, thereby establishing a more accurate image classification model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing image classification models are used for GPR images, then the classification process can be performed, but the classification accuracy is insufficient due to difficulty in extracting features from complex and diverse GPR images
Solution Approach 1:
The patent segments the training process into two distinct phases: primary learning for feature value extraction and secondary learning for image classification. This segmentation allows the model to first master feature extraction from complex GPR images, then apply those features to classification tasks, thereby improving classification accuracy while addressing the inherent difficulty of feature extraction from diverse GPR image patterns
Solution Approach 2:
The patent implements preliminary action by performing primary learning to establish feature value extraction capabilities before conducting secondary learning for classification. During primary learning, the model learns to extract meaningful features from complex and diverse GPR images, creating a solid foundation that enables accurate classification in the subsequent secondary learning phase
2Reliability
If a single unified training process is used for both feature extraction and classification, then the training process is simpler, but the model cannot achieve optimal performance for complex GPR images requiring separate learning stages
Solution Approach 1:
The patent divides the training process into two separate learning stages: primary learning focused on feature value extraction and secondary learning focused on image classification. This segmentation enables the model to achieve optimal performance for complex GPR images by dedicating each stage to its specific learning objective, thereby improving reliability without requiring overly complex training procedures
Solution Approach 2:
The patent introduces feature values as an intermediary between the input GPR images and the final classification output. The primary learning stage learns to extract these intermediary feature values from images, and the secondary learning stage uses these features for classification, creating a structured two-stage process that improves model performance while maintaining manageable training complexity
Data Source
AI summary
A method and system for training an image classification model is disclosed. An aspect is to separate training processes of a feature value extraction model and an image classification model and train the feature value extraction model on a representative feature value suitable for image classification into a specific label value (e.g., “Peak”), thereby improving accuracy and performance of a classification model for a ground-penetrating radar (GPR) image that is captured by a GPR and is not easy for feature value extraction.