Region of Interest (RoI) Detection in Ground Penetrating Radar (GPR) Data 2D ENTROPY ANALYSIS Presenter: Yu Zhang
2 Autobiography RESEARCH INTERESTS
3 Autobiography University of Vermont Doctor of Philosophy (Ph.D.) in Electrical Engineering 2012 – Present Advisor : Dr. Tian Xia Huazhong University of Science and Technology Bachelor of Science (B.S.) in Electrical Engineering 2008 – 2012
4 Research Interests Compressive OFDM GPR 30% Compression Tx Rx Full OFDM Spectrum Rebar Concrete 5% Compression
5 Research Interests Synthetic Aperture Radar (SAR) based GPR imaging Triangle Rectangle Cylinder Test Scenario GPR-SAR B-Scan Image Regular GPR B-Scan Image
6 Research Interests Low-Rank and Sparse Representation in GPR and Through-the-Wall Radar Imaging Regular Clutter Removal Test Scenario GPR B-Scan L + S Representation
7 Research Interests GPR Signal Processing related Problems Image De-noising Clutter Removal Image Migration Region of Interest (RoI) Detection
8 Introduction WHY GPR, GPR OPERATING MECHANISM & OUR GPR SYSTEM
9 Why GPR? Non-destructive evaluation (NDE) of transportation infrastructure. GPR as a highly efficient NDE method: Concrete bridge deck inspection; Asphalt pavement monitoring; Highway rebar detection; Railroad ballast condition assessment, Soil moisture estimation.
10 GPR Operating Mechanism Subsurface medias of different dielectric constants Each position: A-Scan trace Assemble all A-Scan traces: B-Scan image
11 Home made GPR System Ultra-wide band (UWB) pulse generator UWB antenna High speed digitizer configuration Wheel encoder FPGA controller PC & LabVIEW user interface
12 Prototype
Xia & Huston sitting on GPR 13 A ROBUST system package that can hold two adults’ weight!!!
14 Air-coupled Impulse GPR System Specifications Data acquisition unit 8 Gsps, 10-bit resolution Sampling window width 40 ns Pulse repetition frequency 0 to 30 kHz tunable Horizontal resolution 1 cm at 60 miles/h survey speed Signal bandwidth 600 MHz to 2 GHz tunable Penetrating capability Up to 1 meter
15 Today’s Topic ROI DETECTION IN GPR DATA USING 2-D ENTROPY
16 Why 2-D Entropy Large volume (overall 20 miles) railroad GPR data set collected during the field test at Metro St. Louis and Massachusetts Bay Transit Authority. Data collection in St. Louis Forest Park Station to Sunnen Station High resolution GPR system brings us ~300 GB data. How to process such large volume data???
17 Why 2-D Entropy Large data volume Sophisticated data processing is too computationally complex and even infeasible Properties of subsurface scatters or material are too complex Obtain of prior knowledge or training data is unrealistic It is desirable to Develop an unsupervised and automatic GPR data processing method that can effectively and rapidly identify suspicious features from big radargram
18 First Glance at 2-D Entropy Analysis Entropy is a measure of the uncertainty associated with a random variable. Entropy characterization is explored to identify singular regions within a large GPR data set High entropy value indicates high similarity Low entropy value specifies high singularity
19 What do we have in a GPR railroad B-Scan image Noises Clutter Multiple Peaks from one Scattering Cross-ties or Sleepers Useless Background
20 Unsupervised GPR RoI Detection based on Entropy Power Information Subsurface Pre-processing Characterization Identification Raw Stacking Every Low Pass Hilbert A-Scan Data 50 Traces Filter Transform Decomposition Region of 2-D Entropy Stacking Every Clutter Subsurface Interest Analysis 10 Trace Removal Identification RoI Detection B-Scan Image Enhancement
21 Pre-Processing Step 1: Stack every 50 A-scan traces to calculate the average to boost the signal-to-noise ratio (SNR). o The selection of 50 traces for calculation considers the balance between the obtainable image resolution and noise reduction performance Step 2: Apply Low Pass Filtering (LPF) with a 2 GHz cutoff frequency. o GPR transmission signal’s frequency: 600 MHz – 2 GHz
22 Power Information Characterization Hilbert Transform is implemented to extract the pulse envelope. Hilbert Transform of signal 𝑡 𝑢 𝑡 𝑢 = ℋ 𝑡 = ℎ 𝑢 ∗ 𝑡 𝑢 = 1 𝜌𝑢 ∗ 𝑡(𝑢) Analytic signal 𝑡 𝑏 𝑢 = 𝑡 𝑢 + 𝑗 𝑡(𝑢) Signal’s envelope 𝑡 𝑢 2 + 𝑡(𝑢) 2 𝑡 𝑏 𝑢 =
23 Power Information Characterization (con.) Ricker wavelet source – 3 peaks 1 st pulse - antennas’ direct coupling 2 nd pulse - reflection signal from 1 st scatter 3 rd pulse – reflection signal from 2 nd scatter GPR A-Scan trace GPR A-Scan envelope
24 Subsurface Identification A-Scan decomposition is performed to isolate subsurface layer and narrow down the scope of data Transmitter and receiver antennas’ direct coupling pulse as the reference signal By performing iterative cross correlations, an A-Scan waveform is decomposed into the combinations of multiple pulses of varying amplitude and phases characterizing the reflection signals from different scatters
25 Subsurface Identification (con.) 1 st backscattering pulse 2 nd backscattering pulse Direct coupling signal
26 B-Scan Image Enhancement Step 1: Remove the background signal using a 2-D High Pass Filter (HPF) o In horizontal direction, the frequency bandwidth of clutter is much narrower than that of subsurface scattering signals. Step 2: Stack every 10 A-scan traces o Further improve signal SNR as well as reduce data volume and redundancy
Renyi’s Entropy 27 GPR backscattering signal 𝑍 𝑢 can be modeled as 𝑍 𝑢 = 𝐸 𝑢 + 𝑇(𝑢) 𝐸 𝑢 - reflection signal from objects of interest, 𝑇(𝑢) - background signals. Power normalization is first performed 𝑗 (𝑢) 2 𝑍 𝑧 𝑗 𝑢 = 𝑁 𝑗 (𝑢) 2 𝑗=1 𝑍 𝑧 𝑗 𝑢 - normalized signal, 𝑗 - trace index, 𝑁 – total number of traces included, 𝑢 - time index of data points on each trace. 1 𝑁 𝛽 Renyi’s entropy: 𝐹 𝛽 𝑢 = 1−𝛽 log e 𝑗=1 𝑧 𝑗 𝑢 𝛽 – the entropy order. ( 𝛽 = 3 here)
28 2-D Entropy Based RoI Detection Step 1: 2-D Renyi entropy calculation Step 2: Entropy curve smoothing using moving average method: 𝑜 𝐹 𝑡𝑛𝑝𝑝𝑢ℎ 𝑜 = 1 𝐹(𝑗) 𝑛 𝑗=𝑜−𝑛+1 𝐹(𝑗) – entropy value of 𝑗 th data in entropy sequence, 𝑛 is selected as 5% of the total number of entropy data points o Optimal smoothing performance as well as data resolution Step 3: Adaptive entropy threshold determination using OTSU method. o Minimize inner-group variance & maximize inter-group variance
29 Lab Experiment The ballast layer of 0.3 meters thickness is laid above the soil 0.75 meters apart from the left end of the platform, an area of 0.45 meters length and 0.2 meters depth is filled with the fouled ballast, which is a mixture of sand, ballast, and water Side view for the subsurface structure Railroad ballast test platform
30 RoI Detection (1) Raw B-Scan image Pre-processed B-Scan image
31 RoI Detection (2) Signal Envelope Signal Decomposition
32 RoI Detection (3) Ballast layer image Enhanced B-Scan ballast layer image
33 RoI Detection (4) Entropy along Travel Time (y-axis) Entropy along Scan Axis (x-axis)
34 RoI Detection (5) RoI Detection Result based on 2-D Entropy analysis
35 Railroad Ballast Field Test Home made GPR system mounted on vehicle during field tests
36 Test Sites Metro St. Louis MetroLink blue line Railroad ballast (1:00~4:00 am)
37 RoI Detection (1) Raw B-Scan image Pre-processed B-Scan image
38 RoI Detection (2) Signal Envelope Cross-tie marked by signal decomposition
39 RoI Detection (3) Ballast layer image Enhanced B-Scan ballast layer image
40 RoI Detection (4) Entropy along Travel Time (y-axis) Entropy along Scan Axis (x-axis)
41 RoI Detection (5) Suspicious fouled ballast regions marked by white rectangle
42 Conclusions The unsupervised automatic ROI detection method developed in this study can effectively identify regions of interest in subsurface for further in-depth analysis. The proposed unsupervised automatic GPR data processing algorithm has been effectively applied to laboratory and field test data. The analysis results prove that the proposed algorithm can correctly identify the region of interest and can accurately measure the region’s location .
43 Future Work Local entropy & multi-scale entropy analysis
44 Q & A? Q&A?
Recommend
More recommend
