MND-GAN: A Research on Image Deblurring Algorithm Based on Generative Adversarial Network

Aiming at the image blurring problem that often occurs in machine vision applications in intelligent transportation and unmanned driving scenarios, an image blurring algorithm MobileNetDeblur-GAN (MND-GAN) based on generative adversarial network is proposed in this paper. The algorithm has the following characteristics: the least square generative adversarial model (LS-GAN) is used for network training to solve the gradient disappearance problem in the traditional GAN network training process; the generator structure is designed based on MobileNet and two scales of input images are used to improve the authenticity of image restoration at different scales; the weights of nonlinear layers at different scales are shared, which greatly reduces the overall reasoning time of the deblurring algorithm; a global discriminator and a local discriminator are combined to construct a dual discriminator structure to improve the adaptability of the model to different scenarios; the loss function uses the combination of perception loss and structure loss to make the deblurring effect more consistent with the visual characteristics of clear images. The performance of MND-GAN algorithm is compared on multiple datasets, and good deblurring effect is obtained. At the same time, the reasoning speed of the network exceeds that of other comparison algorithms, and it can meet the real-time requirements of intelligent transportation and unmanned driving scenarios.