Multi-Classifier Buried Mine Detection Using MWIR Images

Dr. Bo LingMigma Systems, Inc.

Mr. Anh H. Trang Mr. Chung PhanUS Army RDECOM

April 10, 2007

Defense and Security Symposium 2007

Presentation Outline

- Overall Technical Approach

- MWIR Image Thresholding and Clustering

- Buried Mine Directional Signatures

- Multi-Classifier for Buried Mine Detection

- Test Results

- Conclusion

Overall Approach

Image Thresholding

Image Clustering

Image Windowing

Buried Mine Classifier

(horizontal)

Buried MineImage

Buried Mine Classifier(vertical)

Buried Mine Classifier(diagonal)

Fusion

Image Thresholding Using Wavelet Transform

Image Thresholding Based on Inverse Wavelet Transform

imagei

imageiii YifT

YifYY

ˆ

dvhWimage tttf ,,ˆ1

where is related to the inverse of discrete wavelet transform, th, tv , and td are the threshold values associated with three decompositions in the wavelet domain.

1Wf

Image Thresholding

100 200 300 400 500 600

50

100

150

200

250

300

350

400

450

500

Original Image Thresholded Image

Thresholding method has preserved the surface and buried mines.

Image Clustering

Image Thresholding

Image Clustering

Image Windowing

Buried Mine Classifier

(horizontal)

Buried MineImage

Buried Mine Classifier(vertical)

Buried Mine Classifier(diagonal)

Fusion

Adaptive Self-Organizing Maps (ASOM)

Data

No prior knowledge of number of clusters

neurons

Neuron activation function

Similarity Measurement

2)(2

),( ic

ii eawx

wx

j j jj

k

j jji

i

i

dd

ddc

22

1),(wx

wxwx

Clustering after Thresholding

Clustering

100 200 300 400 500 600

50

100

150

200

250

300

350

400

450

500100 200 300 400 500 600

50

100

150

200

250

300

350

400

450

500

ClustersThresholded Image

Each cluster represents a potential mine

Buried Mine Signatures

The similarity-based 3D ASOM is used to find clusters in the windowed target chip.

Original ImageTarget Chip Clusters

Directional Scanning

We build buried mine signatures in three directions

Horizontal Scan Vertical Scan Diagonal Scan

Library of Buried Mine Signatures

We have found that the thermal variation patterns exhibited in daytime and nighttime are significantly different.

Signature Vectors

Horizontal Signatures Vertical Signature Diagonal Signature

Example of Buried Mine Signatures

Target Chip

Signature

Histogram

The signatures associated with buried mines are common in

- Long vector length

- Histogram peaked in the middle

Signature Comparison

Mine Signatures

False Alarm Mitigation

Signature difference can be used to eliminate false alarms.

Multi-Classifier Detection

Image Thresholding

Image Clustering

Image Windowing

Buried Mine Classifier

(horizontal)

Buried MineImage

Buried Mine Classifier(vertical)

Buried Mine Classifier(diagonal)

Fusion

Three Directional Classifiers

Horizontal Classifier

Vertical Classifier

Diagonal Classifier Each of three

classifiers will process the corresponding directional signatures.

Test Result of Nighttime Image

We have tested both daytime and nighttime images taken from MWIR data collected as part of Lightweight Airborne Multispectral Minefield Detection (LAMD) program.

Original Image

Test Result - Clustering

Original Image Clustered Image

Since each cluster could represent a buried mine, we must process all clusters.

Test Result - Three Classifiers

Each cluster is windowed and processed by all three directional classifiers.

There are three independent detection results.

Three false alarms Three false alarms Four false alarms

Test Result - Fusion

We have used a simple fusion scheme: a buried mine is declared only if it is detected by all three classifiers.

One advantage of this type of fusion is low false alarm rate since three classifiers may not report the same false detection in the same image.

Final Detection

Two false alarms left.

They can be further eliminated.

Conclusion

For each target chip, we scan it in three directions: vertical, horizontal, and diagonal to construct three signatures.

For the same target chip, there will be a total of three classifiers associated with vertical, horizontal, and diagonal scans.

These three classifiers are applied to the same target chip, resulting in three independent detection results, which are further fused for a refined detection.

New results will be reported in the future once we test the system with new images.