Image Feature Classification via Saliency and Mode Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image processing techniques face challenges in effectively classifying features in medical images, particularly in capturing informative regions without scene-specific learning, which limits their versatility and accuracy in identifying normal or abnormal medical conditions.
Innovation Solution
A method involving computing a saliency surface, conducting a mean shift algorithm to generate mean shift data, producing a mode voting map, and classifying features using this map, which includes determining distinctiveness scores and areas of influence for each mode, allowing for the application of medical condition classifiers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If saliency algorithms are used to capture informative portions of images, then versatility is improved, but classification accuracy for medical conditions deteriorates
Solution Approach 1:
The image is divided into multiple patches, and a saliency surface is computed for each patch independently. This segmentation allows the algorithm to capture local informative regions while maintaining overall versatility across different image types, resolving the contradiction between versatility and classification accuracy.
Solution Approach 2:
A mode voting map is introduced as an intermediary data structure between the saliency surface and the final classification. The mean shift algorithm processes the saliency surface to generate this intermediate representation, which then feeds into the medical condition classifier. This intermediary layer enables both versatile saliency detection and accurate medical condition classification.
2Measurement precision
If detailed saliency data is processed through mean shift algorithm, then classification accuracy is improved, but data complexity increases
Solution Approach 1:
The mean shift algorithm extracts only the essential mode information from the detailed saliency surface data. By identifying and isolating the dominant modes and their areas of influence, the algorithm reduces the complex saliency data into a simplified mode voting map that retains classification accuracy while reducing data complexity.
Solution Approach 2:
The transformation from saliency surface to mode voting map involves changing the data representation parameters. The continuous saliency values are converted into discrete mode labels with associated voting weights, fundamentally altering the data structure to reduce complexity while preserving the essential information needed for accurate classification.
3Productivity
If mode voting map is produced from mean shift data, then classification speed is improved, but information loss increases
Solution Approach 1:
The mode voting map incorporates feedback from the mean shift analysis by assigning voting weights to different modes based on the area of influence calculations. This feedback mechanism ensures that modes with larger areas of influence contribute more to the final classification, preserving important information while enabling faster processing through the simplified map structure.
Data Source
AI summary
A method, executed by one or more processors, includes computing a saliency surface for an image, conducting a mean shift algorithm on the saliency surface to provide mean shift data for the saliency surface, producing a mode voting map for the saliency surface from the mean shift data, and classifying features in the image according to the mode voting map. The features may correspond to medical conditions. In some embodiments, computing the saliency surface comprises determining a distinctiveness score for each of a plurality of image patches. In some embodiments, producing the mode voting map comprises determining a plurality of modes and an area of influence for each mode of the plurality of modes where the area of influence corresponds to mean shift data that leads to a particular mode. A corresponding computer program product and computer system are also disclosed herein.