Positron Image Super-Resolution Using Generative Adversarial Networks

Abstract

Positron images generated by positron non-destructive testing technology under rapid detection scenes such as low concentration dose, low exposure time and short imaging time, which have some problems like low-resolution and poor definition. These issues cannot be solved for the being time. To solves these problems, this research super-resolves the low-resolution positron images to generate images with high-resolution and clear details. To make the generated super-resolution images more capable of restoring the features of low- resolution images, this research proposed a positron image super-resolution reconstruction method based on generative adversarial networks. In order to improve the input information utilization rate, long skip connections were added into the generator. In addition, the discriminant model, where composed of an image discriminator and a feature discriminator, can stimulate the generator to generate clearer super-resolution images which contain more details. In attempting to solve the problem of dataset matching, a special positron image super-resolution dataset is constructed for network application scenarios. In the adversarial training stage, perceptual similarity loss and adversarial loss are used to replace the traditional mean squared error loss to improve the images perception quality. Experimental results show that the proposed model can reconstruct low-resolution images by four times super-resolution in 0.16 seconds. The super-resolution images obtained are superior to other algorithms in visual effect, which have clearer detail structure and higher objective performance values. Hence this model can meet the requirements of rapid non-destructive testing of industrial parts.