Inpainting algorithm based on hybrid dilated convolution and improved knowledge consistent attention

In order to effectively solve the problems of structural loss and local texture distortion in repairing large-scale and continuous holes in an image，an image restoration algorithm based on the improved knowledge consistent attention mechanism is proposed． Firstly，the partial convolution is performed to normalize the corrupted image and update the mask． Subsequently，the result is fed into the region identification module composed of two partial convolutions． Thereafter，the recognized information maps are sent into the feature induce structure containing the hybrid dilated convolution (HDC) and the improved knowledge consistent attention (KCA) mechanism． The features of the output images are combined after the induce iteration is completed．Finally，the reconstructed feature map is subjected to the post-processing by using deconvolution and de-residual bottleneck network to enhance the structural integrity of the repaired image．Group normalization (GN) are applied in the proposed module improve the convergence speed of the loss function． Experiments are fulfilled on public data set of Paris street view． The results demonstrate that the proposed algorithm can effectively inpaint large-scale and continuous damage in an image and avoid restoration distortion．Moreover，the proposed method demonstrates superior performance on the peak signal-to-noise ratio，structural similarity and time consuming compared with the state-of-the-art methods. © 2023 Huazhong University of Science and Technology. All rights reserved.