基于改进Faster R-CNN算法的太赫兹安检图像识别检测

Recognition and Detection of Terahertz Security Images Based on an Improved Faster R-CNN Network Algorithm

太赫兹成像系统产生的太赫兹安检图像存在着色调单一、样本单一、数据量少、分辨率低、清晰度和对比度差等特点,针对上述问题,通过DC-GAN网络生成了更多的太赫兹安检图像数据集,并采用ESRGAN网络对太赫兹安检图像进行了超分辨重建以及线性变化阈值处理,使得图像细节纹理特征更清晰,并滤去了大量的背景噪声使得图像的对比度得到了增强。同时,针对太赫兹安检图像中检测目标(刀、手机)存在较多重叠的情形,对网络中非极大值抑制算法存在的不足进行了改进。引入Sigmoid加权的方法避免了与目标重叠较大的检测框被直接删除,通过降低其置信度,解决太赫兹安检图像中由于目标堆叠造成漏检的问题。实验结果验证了改进的Faster R-CNN网络对太赫兹安检图像可疑物体检测拥有更高的准确性。

The terahertz image produced by terahertz imaging system has the characteristics of single hue and sample,low resolution,sharpness,and contrast,and small amount of data.In order to solve these problems,more terahertz image data sets were generated through the DC-GAN network,and the ESRGAN network was used to accomplish terahertz image super-reconstruction and linear transform threshold processing,which makes the image details and texture features clearer,and filters out a lot of background noise to enhance the image contrast.At the same time,the non-maximum suppression algorithm in the network was improved for the situation where there are many overlapping detection targets(knives,mobile phones)in the terahertz security image.The Sigmoid weighting method was introduced to avoid the detection frame with a large overlap with the target being directly deleted.By reducing its confidence,the problem of missed detection due to target stacking in the terahertz security inspection image was solved.The experimental results verify that the improved Faster R-CNN network has higher accuracy for detecting suspicious objects in THz security image.