Abstract
As an essential method for security inspection in nuclear facilities, digital radiography has the ability to find hidden contraband efficiently. However, the images obtained by current scanning digital radiography system can be degraded by several factors, such as statistical noise and response time of detectors. At high scanning speed, the statistical noise and vibration of the system deteriorates the quality of images. In addition, the reduction of image quality will influence the accuracy of image observation and recognition. To meet the demand of detection efficiency and quality, it is necessary to guarantee the quality of images under high scanning speed.
Thus, to improve image quality of vehicles’ digital radiography at a certain scanning speed, we proposed an approach (VDR-CNN) to reduce or eliminate image noise, which is a convolutional neural network (CNN) with residual learning. The high-quality images obtained at low scanning speed of system served as the ground-truth image for VDR-CNN, while the low-quality counterpart corresponding to the high scanning speed served as the input. Then, the two images mentioned above constitute a training pair. By training this network with a set of training pairs, the mapping function of promoting image quality will be automatically learned so that the restored image can be obtained from the low-quality counterpart through the trained VDR-CNN. Moreover, this method avoids the difficulty in figuring and analyzing the complicated image degradation model.
A series of experiments was carried out through the 60Co inspection system developed by Institute of Nuclear and New Energy Technology, Tsinghua University. The experimental result shows that this method has attained a satisfying result in denoising and preserving details of images and outperforms BM3D algorithm in terms of both image quality improvement and the processing speed. In conclusion, the proposed method improves the image quality of vehicles’ digital radiography and it is proved better than traditional methods.
Automated Identification of Wood Veneer Surface Defects Using Faster Region-Based Convolutional Neural Network with Data Augmentation and Transfer Learning
