Research On Image Inpainting Technology Based On Attention Mechanism

Over the past two decades,image inpainting has been extensively studied in the field of image processing.It aims to fill in missing or corrupted parts of an image with satisfactory and reasonable content.Traditional techniques have limitations when dealing with large or complex missing regions,as they struggle to generate semantically compliant images.Recent advancements in deep learning and adversarial learning have shown promising results in image inpainting.However,existing methods may still produce distorted structures and hazy textures for large missing regions.This is primarily due to their limited consideration of global or long-range structural information caused by locality of vanilla convolution operations,even with dilated convolutions.To overcome this issue,we first propose a new image inpainting network called the Multi-Scale Self-Attention Generative Adversarial Network(MSSA-GAN),which can obtain feature information at different scales.Specifically,MSSA-GAN including some new Cascaded Self-Attention Propagation Module at middle and depth layer,it can help MSSA-GAN to focus on structural information and textures at different scales to cope with image inpainting tasks at different scales.Moreover,we propose a novel image inpainting network called the Hybrid Dual Attention Generative Adversarial Network(HDA-GAN),which allows capturing both global structural information and local detailed textures.Specifically,HDA-GAN integrates two types of cascaded attention propagation modules,namely,Cascaded Channel-Attention Propagation and Cascaded Self-Attention Propagation,into different convolutional layers of the generator network.For the Cascaded Channel-Attention Propagation module,we concatenate several Multi-Scale Channel-Attention Block into shallow layers to learn features from low-level details to high-level semantics.The Multi-Scale ChannelAttention Block adopts the split-attention-merge strategy and residual-gated operations.Through the split-attention-merge strategy and residual-gated operations,the MultiScale Channel-Attention Block is able to aggregated multiple channel attention correlations for enhancing high-level semantics while preserving low-level details.For the Cascaded Self-Attention Propagation module,we stack several PositionalSeparated Self-Attention Block into middle and deep layers.The Positional-Separated Self-Attention Block also adopts the same split-attention-merge strategy and residualgated operations as the Multi-Scale Channel-Attention Block,but with some changes.The purpose of this design is to use the Positional-Separated Self-Attention Block maintain the details better and while learning long-range semantic information interaction.Besides,the design of the Positional-Separated Self-Attention Block effectively reduces the computational complexity compared with original self-attention.Tests conducted using the Paris Street View and Celeb A-HQ datasets demonstrate that HDA-GAN outperforms several state-of-the-art algorithms in terms of image inpainting quality and quantity.The images were resized to 512x512 or 256x256,following standard settings for training and testing.Evaluation metrics such as Mean Square Error(MSE),Peak Signal-to-Noise Ratio(PSNR)and the Structural Similarity Index(SSIM)were used to assess the performance of different methods.Comparisons were made with various approaches,considering different hole-to-image region ratios.In the experiments with the Paris Street View dataset,the proposed method achieved higher PSNR and SSIM compared to the Edge-LBAM method.Similarly,in the Celeb A-HQ datasets experiments,the proposed method demonstrated lower MSE values and improved PSNR compared to the AOT-GAN method.The results not only showed quantitative improvements but also significant qualitative enhancements.The study introduced a novel hybrid attention generative adversarial network called HDAGAN,which effectively addresses complex missing regions and large holes by incorporating cascaded attention propagation modules.This approach improves the capture of global structure and local texture,resulting in superior inpainting results.