Images’ Partially Blurred Part Location and Restoration based on the Calculation Model and GAN Algorithm

Abstract

Digital images would be blurred due to defocus and motion of objects. A lot of researches have been done on the restoration of motion blurred images. Unlike the global motion blurred image, the local blurred image needs different methods to recover as there is both clear and blurred part in an image. In this paper, we proposed a method combining calculation model and Generative Adversarial Network (GAN), which can automatically identify the local blurred region and deblurring. Also, the blurred part can be filled back correctly. Firstly, the gradient image of the complete image is calculated based on the Sobel operator. The boundary information of the fuzzy region and the Gauss function variance of the blurred and clear region are obtained. And the power spectral gradient of the whole image and each local pixel block are further compared with the variance of the Gauss function. After the analysis is accomplished, a more accurate fixed position is achieved. After finishing extracting with clipping pixels, the extracted fuzzy region is input into the pre-trained Generative Adversarial Network to restore the local image. Finally, the alpha channel algorithm is used to calculate the RGB component without the alpha channel of the two images. The simulation results show the feasibility of this method.