论文解读《Deep Convolutional Neural Networks for Classifying GPR B-Scans》

标题：Deep Convolutional Neural Networks for Classifying GPR B-Scans
作者：Lance E. Besaw and Philip J. Stimac
期刊：Proc. of SPIE Vol. 9454 945413-1

abstract

Symmetric and asymmetric buried explosive hazards (BEHs) present real, persistent, deadly threats on the modern battlefield. Current approaches to mitigate these threats rely on highly trained operatives to reliably detect BEHs with reasonable false alarm rates using handheld Ground Penetrating Radar (GPR) and metal detectors. As computers become smaller, faster and more efficient, there exists greater potential for automated threat detection based on state-of-the-art machine learning approaches, reducing the burden on the field operatives.
Recent advancements in machine learning, specifically deep learning artificial neural networks, have led to significantly improved performance in pattern recognition tasks, such as object classification in digital images. Deep convolutional neural networks (CNNs) are used in this work to extract meaningful signatures from 2-dimensional (2-D) GPR B-scans and classify threats. The CNNs skip the traditional “feature engineering” step often associated with machine learning, and instead learn the feature representations directly from the 2-D data. A multi-antennae, handheld GPR with centimeter-accurate positioning data was used to collect shallow subsurface data over prepared lanes containing a wide range of BEHs. Several heuristics were used to prevent over-training, including cross validation, network weight regularization, and “dropout.” Our results show that CNNs can extract meaningful features and accurately classify complex signatures contained in GPR B-scans, complementing existing GPR feature extraction and classification techniques.

掩埋爆炸危险（BEH）在现代战场上有着致命的威胁。随着计算机的微型化和机器学习的快速发展，快速和智能的监测成为可能。本文结果显示，CNN能够提取有效特征，对信息量丰富的b-scan图像可以正确分类。

introduction

Buried explosive hazards (BEHs) are dangerous and persistent threats that have, and will continue to impact, worldwide areas of conflict for decades.1,2 BEHs can include landmine threats; deeply buried metallic and non-metallic improvised explosive device (IED) main charges; and triggering mechanisms to include non-metallic conductive components and wires. These symmetrical and asymmetrical explosive hazards are contained in a variety of materials (e.g., metal, plastic, wood, glass) which makes their detection and neutralization very difficult. Recent advancements in electronics, radio antenna design and digital signal processing have allowed GPR to excel beyond other sensing modalities in detecting BEH. GPR is a non-destructive geophysical technique used for a variety of shallow subsurface imaging applications. For BEH detection, a GPR transmit antenna directs a pulse of radio frequency (RF) radiation into the ground and a receiving antenna measures the return signal’s amplitude as a function of time (or depth) – called an A-scan. The structure of an A-scan is strongly affected by the medium through which the radiation propagates. If the medium contains regions with different dielectric constants, the A-scan will exhibit complex reflections at the region interfaces. Individual A-scans characterize a narrow “slice” of the subsurface. B-scans are generated by collecting many A-scans and assembling them in spatial sequence to provide a more effective means for visualizing and characterizing the shallow subsurface.

*上面有一段关于a-scan, b-scan的描述

2 experimental data

The data used in this research were collected using a developmental Niitek handheld GPR sensor. The detector has multiple unique transmit-receive pairs that produce multiple channels of data. Data were collected at an Government Eastern Test Range (ETR) using a robot to precisely swing the handheld GPR detector laterally and reposition the GPR detector down lane after each swing. This system was used to interrogate the shallow subsurface over several prepared lanes containing various types of BEHs at various depths. ==The robot was programmed to sweep across the lanes with a particular velocity profile that produced consistent data across the lane. == At the end of each swing, the robot drove down the lane a defined distance and swung the detector back across the lane in the opposite direction. Throughout the process, the position of the detector was measured in real-time using GPS and kinematic sensors, thus generating accurate 3D GPR maps of the subsurface. Representative B-scans over buried anti-tank (AT) and anti-personnel (AP) landmine targets are shown in Figure 1. For the AT target (left panel) we see large “disturbance” (reflection) and a hyperbolic signature spanning more than 100 packets (individual A-scans) of GPR data. For the much smaller AP mine (right panel) we still see a fairly strong reflection, but the spatial extent is less than 100 packets due to the smaller target profile. The high amplitude (white) band occurring roughly at depth bin 40 in both images is the point at which the radiated RF waves are reflected from the air-ground interface. Over very uneven or rough terrain, this “ground bounce” can complicate the interpretation of GPR data, leading to missed targets and false alarms close to the ground surface. In our processing pipeline, we take special care to track and mitigate effects of the ground bounce.

4 ALGORITHMS

Our overall GPR processing pipeline consists of many components found in traditional GPR software, including preprocessing, prescreening (anomaly detection), post processing, discrimination and algorithm fusion. Our preprocessing steps include spatial resampling, ground-bounce tracking and alignment, and A-scan phase alignment. We then use a Deep Belief Network as a pre screener and a series of postprocessing steps to identify anomalous signatures. GPR B-scans are extracted from each of these anomalies and used to train and evaluate our BEH discrimination algorithms. More details of many of these GPR processing pipeline components can be found in reference 17. After extracting the B-scans, we apply a 2D median filter over them as well as a zeros scores component analysis (ZCA) technique to smooth and whiten the GPR B-scans. An easily detected AP target is presented in Figure 3 at various stages of our processing pipeline. Figure 3a shows the raw GPR B-scan (same as right target in Figure 1). Figure 3b shows the same target after ground-bounce tracking and alignment as well as phase alignment and linear depth scaling. Figure 3c shows the target signature after ZCA whitening. ZCA has effectively removed many systematic features evident in Figure 3b, especially the dark horizontal bands without severely altering the fundamental properties of the image, as is often observed with principal component analysis whitening. ==Finally, Figure 3d shows the 49x49 “pixel” GPR B-scan that is used as input to the CNN discrimination algorithm. ==

经过预处理和噪声去除，最终切割获得49x49的b-scan图像。CNN算法将这样尺寸的图像作为输入。

The above pipeline was used to detect anomalies using data collected from the ETR in 2013 and 2014. More than 1,800 m2 of GPR data was used in this study, including a total of 786 unique target encounters. With detection thresholds set arbitrarily low, to ensure all of the BEH were detected, our processing pipeline detected all BEH targets but also recorded many FAs. We extracted the GPR B-scans (49x49 pixel image) form each anomaly from the closest GPR channel and used it as input to our traditional feature processing and CNN algorithms.
Packet

As a means for feature extraction and discriminant function comparison, we employ two common methods in the BEH discrimination literature.

discriminant function comparison：判别函数比较

As shown in Figure 2, the CNN used in this work consisted of two convolutional layers, each with max-pooling layers, followed by a single layer of fully connected hidden nodes. We used the tanh activation function in both the convolutional and hidden layers. A total of 20 filters were employed in each convolutional layer and 100 nodes in the fully connected hidden layer. We also used the technique known as “dropout” when training the hidden layers with a dropout rate of 0.5.16 As with other CNN applications, including dropout in training provided significant improvement in training (2x classification accuracy). We performed max pooling after each convolutional layer with a 2x2 pooling mask. Following the hidden layer, we used a 2-way softmax classification layer to associate each input image as either a BEH or FA. These hyperparameters were used based on some grid search over relatively small ranges. We did not experiment with deeper CNNs (i.e., more convolutional layers and/or more fully connected layers). As with the other algorithms, we used 3-fold cross validation to train and evaluate the CNN performance in discriminating BEH from FAs. Our CNN implementation allowed us to speed up its training using Graphical Processing Units (GPUs). Our entire network and training data easily fit on a single NVIDIA C2075 Tesla GPU. This GPU has 448 processing cores, running at 1.15GHz and 6GB of memory and is capable of 515 giga floating point calculations per second (Gflops). Using this GPU for CNN training we were able to reduce training time from ~16 hours on the CPU (2.6GHz Intel Xeon processor) to just 1 hour (16X speedup). The CNN and its optimization algorithm were written in Python and its deep learning extension, Theano.18

总结

本文用cnn网络通过输入的49*49的b-scan图像及分类，学习分类算法后，输出样本的分类（为beh,fa两类）。