红外与毫米波学报, 2017, 36 (3): 354, 网络出版: 2017-07-05
利用卷积神经网络进行毫米波图像违禁物体定位
Using convolutional neural network to localize forbidden object in millimeter-wave image
毫米波图像 卷积神经网络 物体定位 概率累积图 millimeter-wave image convolution neural network(CNN) object localization probability accumulation map
摘要
随着毫米波器件的成熟, 毫米波成像雷达已经应用于人体安检.但毫米波图像中违禁物体的定位仍然是一个艰巨的任务, 这极大地限制了毫米波成像雷达的应用.文章将卷积神经网络(Convolutional Neural Network, CNN)应用于毫米波图像, 自动定位毫米波图像中的违禁物体, 如枪、刀等.利用滑动窗口在输入图像上滑动, 并通过CNN得到各个子图块存在违禁物体的概率.图像块是相互交叠的, 将各子图块的概率值累积起来, 得到概率累积图.概率累计图反映了违禁物体的位置.由于CNN和概率累积图的应用, 在实验中, 该方法获得了很高的定位准确率, 验证了该方法的有效性.
Abstract
With the maturity of millimeter-wave devices, millimeter-wave imaging radar has been applied to human security check. However, the localization of forbidden objects in millimeter-wave images is still a difficult task, which greatly limits the application of millimeter-wave imaging radar. This paper adopts convolution neural network (CNN) to automatically localize forbidden objects, such as guns and knives, in millimeter-wave images. A sliding window is applied to slide over the input image. Then the probability of the existence of forbidden object in the image patch can be obtained via CNN. The image patches are overlapped with each other, and the probability values of all image patches are accumulated to obtain the probability accumulation map (PA-map). The PA-map reflects the position of forbidden objects. Due to the application of CNN and PA-map, this method achieves a high accuracy of localization in the experiment, which verifies the effectiveness of this method.
姚家雄, 杨明辉, 朱玉琨, 吴亮, 孙晓玮. 利用卷积神经网络进行毫米波图像违禁物体定位[J]. 红外与毫米波学报, 2017, 36(3): 354. YAO Jia-Xiong, YANG Ming-Hui, ZHU Yu-Kun, WU Liang, SUN Xiao-Wei. Using convolutional neural network to localize forbidden object in millimeter-wave image[J]. Journal of Infrared and Millimeter Waves, 2017, 36(3): 354.