An Image Completion Method Based on Stacked Generative Adversarial Networks

Abstract

The invention relates to an image completion method based on a stacked generation confrontation network. The network structure is mainly composed of three stacked generator network layers. First, the mask image is cut into multiple image blocks so that the network can extract the features of different image blocks; then, the present invention puts the completed multi-image block results into the generator of the next layer to further complete image; finally, the completion results of different blocks are applied to a whole mask image to obtain the final completion output. From coarse to fine completion, make full use of the high-level semantic information extracted by the convolutional neural network. And through the image block discriminator to distinguish the real and fake of the generated image and the original image. Experimental results show that the method of the present invention can generate high-quality completion results for images with irregular masks, and the completion results are closer to the original image.

Description

technical field [0001] The present invention relates to the field of image processing, in particular to an image complement method based on a stacked generation confrontation network. Background technique [0002] Recently, image completion tasks based on deep learning methods have been greatly developed, and their application scope has gradually expanded. Image completion is a basic task in the field of image processing. Its difficulty lies in the need to fill in the missing regions with real, natural and semantically correct content. The early image completion algorithm used the nearest neighbor search method to search for the most similar image block in the background area to fill the missing area, but this cannot obtain high-level semantic information of the image, so its completion method cannot produce meaningful content. Other image completion algorithms are based on the goal of learning the distribution of the entire data set, and construct missing content through t...

AI Technical Summary

Problems solved by technology

[0006] 2. The training time of the network and the space consumption of computing resources are large

[0008] 3. The network does not take into account the loss of local detail information when extracting image features

[0009] Existing image completion methods ignore or do not fully consider that a large amount of local detail information will be lost when the image is down-sampled, which will cause the network to output blurred completion results

Some networks may consider the loss of details and adopt a residual block or a residual block-like network structure, but the use of detailed information is not sufficient, and the final output result will still have fuzzy defects.

Images