Research On Image Inpainting Methods Based On Attention Mechanism

Image inpainting,as an important research area in the field of computer vision,has been receiving attention in recent years.With the rapid development of the information age,image inpainting technology has been widely used in cultural,communication,military and other fields.Therefore,the research on image inpainting has gradually become a research hotspot in computer vision,and has important research significance.The goal of image inpainting is to reconstruct and restore images with missing regions,making them more realistic and natural in visual observation.With the development of deep learning technology,image inpainting based on deep learning can reconstruct damaged input images through inpainting models.The current mainstream image inpainting method is to use a generative adversarial network,using the game between the generator and the discriminator to enhance the inpainting ability of the generator.Aiming at the problems of insufficient utilization of feature information and insufficient attention to important features in image inpainting,this paper proposes an image inpainting model based on coherent semantic attention and an image inpainting model based on efficient channel attention,using gated convolution,coherent semantic attention,enhanced gated convolution and efficient channel attention under the generation adversarial network.The specific research contents are as follows.Aiming at the problems of color difference and structural distortion caused by insufficient consideration of the continuity and semantic correlation of image features in the region to be repaired,an image inpainting model based on coherent semantic attention is proposed.Based on the coarse-to-fine network architecture,considering that ordinary convolution applies the same filtering blocks to valid pixels,invalid pixels,and synthesized pixels,and can cause color differences when dealing with irregular masks,gated convolution residual blocks are used to improve.In the refinement network,the coherent semantic attention mechanism is adopted,which not only focuses on the correlation between the region to be repaired and the known region,but also focuses on the continuity of features within the region to be repaired.Compare the inpainting results of the model and other image inpainting models on the public dataset Celeb A-HQ and Places2.Experimental results show that replacing ordinary convolution with gated convolutional residual blocks distinguishes between valid and invalid pixels,and using coherent semantic attention enhances attention to semantic coherence within the region to be repaired,improving the quality of image inpainting results.To address the problem of insufficient use of contextual information resulting in blurring and color shift issues,this article proposes an image restoration model based on efficient channel attention.In a coarse-to-fine network architecture,enhanced gated convolution is used to preserve the gated information under each layer of convolution,and efficient channel attention is used to focus attention on the relationships between key channels,improving the ability to obtain contextual attention from distant spatial locations.Comparative experiments were conducted on the public dataset Celeb A-HQ and Places2.Experimental results show that using efficient channel attention and enhanced gated convolution can improve the quality of the image inpainting results.