Improving image inpainting performance with dual-stage adversarial structure and color-aware networks

Image inpainting has achieved impressive progress in repairing the degraded images having missing regions. The image inpainting process includes reconstructing the degraded images and filling the regions’ absence of color information or data. The existing image inpainting approaches do not apply the known information completely in reconstructing the damaged images, which leads to poor inpainting outcomes. To solve this issue, this research proposes a novel dual-stage adversarial structure and color-aware image inpainting network (DASCAIIN) method for image inpainting. This method includes two modules, such as a structure-aware generator, and a color-aware generator. This research utilizes diverse datasets, such as COCO, CelebA, Places-2_MIT, Paris StreetView, and ImageNet datasets to collect the input images. The corrupted images are inputs to the Sobel operator for generating the gradient map. The encoder–decoder-based structure-aware generator is applied to identify the plausible structural contours at the damaged regions by generating the structural prediction map. The encoder–decoder-based color-aware generator is employed to integrate the texture information at the closed region surrounded by structures. This module generates the inpainting image similar to the original image. The proposed DASCAIIN method applies the membrane computing principles including compartmentalization, hierarchical structuring, and communication to improve the process of creating structure and color-aware features. The feature patch discriminator is utilized to identify the inpainting as well as the ground truth images’ every patch as true or fake. The experiments are conducted comprehensively in terms of various evaluation measures. The results exhibited that the proposed method attained a higher PSNR of 36.1 dB, SSIM of 0.956, and a lower computational time of 0.039 s, respectively. Overall, the proposed method effectively reconstructed the degraded images having incomplete regions with better performance outcomes.