A Wrapper-Based Approach to Image Segmentation and Classification Michael E. Farmer, Member, IEEE, and Anil K. Jain, Fellow, IEEE
大綱 Introduction Introduction Overview of the approach Overview of the approach Experiment: Vision-Base airbag suppression Experiment: Vision-Base airbag suppression application application Experimental result Experimental result
Introduction
Traditional processing The traditional processing flow for image- based pattern recognition consists of image segmentation followed by classification.
Three limitations of traditional processing The object of interest “ should be uniform and homogeneous with respect to some characteristic ” and “ adjacent regions should be differing significantly ” There are few metrics available for evaluating segmentation algorithms Inability to adapt to real-world changes
The contributions in this paper Developing a closed-loop framework for image segmentation to find the best segmentation for a given class of objects by using the shape of the object for classification of the segmented object Using the probability of correct classification of the object to provide an “ objective evaluation of segmented outputs ” The system can adapt to “ real-world changes. ”
Overview of the approach
Wrapper-Based Approach Wrap the segmentation and the classification together, and use the classifier as the metric for selecting the best segmentation. Using the classifier to intelligently re-assemble to solve over-segmented problem. The classification is correct when the minimum distance between the classification of the candidate segmentation and one of the desired pattern classes < T
Traditional vs Wrapper-Base
Experiment: Vision-Base airbag suppression application
Problem Infantor Adult
Challenges Nonuniform illumination Poor image contrast Shadows and highlights Occlusions Sensor noise Background clutter
Variability for the infant class
Proposed approach
Preliminary Segmentation Reduce the number of blobs that must be processed. Once the correlation value for each region is determined, an adaptive threshold is applied, and any region that falls below the threshold is considered a part of the foreground.
Preliminary Segmentation
RegionLabeling Region Labeling Using the EM algorithm with a fixed number of components, and then rely on the classification accuracy to determine if more components are required. Merging the very small blobs by mode filter Merging any regions that are smaller then 20 pixels in size with their larger neighbors
RegionLabeling Results Region Labeling Results
Blob Combiner We have framed the blob combiner problem as one of blob selection, where there exists a subset of blobs that will provide the highest classification accuracy for a given pattern class Forward selection mode Forward selection mode Backward selection mode Backward selection mode
Blob Combiner ( Blob Combiner ( plus-L, minus-R algorithm )
Feature Extraction
Acceleration Methods for Feature Extraction Acceleration Methods for Feature Extraction: Precompute the moments for each blob Compute the moments using only the local neighborhood of each blob. Attain over a ten thousand-fold reduction in processing for each moment calculated.
Classification of Blob Combinations Using the nearest neighbor classifier to compute classification distance Feature 1 Feature 2 : class - A points : class - B points : point with unknown class Circle of 1 - nearest neighbor The point is class B via 1-NNR.
Proposed approach
Demonstrating
EXPERIMENTAL RESULTS
Correct segmentations
Incorrect segmentation