| In recent years,image inpainting technology has become one of the important research directions in the field of computer vision,attracting increasing attention from experts and scholars in various fields.It has been widely applied in the restoration of old photos,object removal,image completion,and other applications.The image inpainting studied in this article refers to the use of image inpainting algorithms to restore damaged or missing areas of a given damaged image,filling in the gaps and completing the image to make it as close to the original as possible,which is a kind of image processing technology.Currently,deep learning-based methods have become the mainstream for image inpainting,among which Generative Adversarial Networks(GANs)with powerful feature representation and learning ability have demonstrated outstanding performance in the field of image inpainting.Compared to traditional image inpainting methods,deep learningbased image inpainting methods can better handle complex damaged areas and generate more natural-looking images.Although most existing GAN-based image inpainting methods typically produce visually convincing results,when the image has large damaged areas or complex background information in the damaged areas,the generated results may suffer from significant artifacts,inconsistent color transitions,and blurry texture details.To address these issues,this paper conducts a detailed analysis of the current research status of image inpainting techniques at home and abroad,and proposes a novel Residual Feature Attention Network(RFA-Net)based on GAN for image inpainting.This model improves the traditional encoder-decoder image inpainting network by introducing a backbone network with texture perception capability,and adopts a residual non-pooling design structure to preserve shallow texture features.It also adaptively enhances the weights of channel and position features that contribute to image inpainting,achieving more refined inpainting of texture details in damaged images.The main contributions and innovations of this paper are as follows:(1)This paper designs and implements an image inpainting model RFA-Net based on residual attention network to address the issues of texture detail blurring and uneven distribution of known region key information in damaged images.The network adopts an encoder-decoder architecture,consisting of three main parts: Residual Attention(RA)module,Multi-scale Feature Enhancement(MFE)module,and Dense Feature Fusion(DFF)module.In the encoder,the RA module combines the technical advantages of residual structure and attention mechanism to focus on extracting image inpainting-related features and achieving fine texture inpainting by retaining shallow image features.In the decoder,the MFE module realizes the concatenation of multi-scale convolution kernel features to enhance the network’s feature extraction ability.To achieve large-scale damaged area inpainting,the DFF module is added after the MFE module to expand the network’s receptive field without increasing the network parameters.(2)The design of the discriminator in this paper is based on the improvement of the Relative Average GAN(Ra GAN),and a multi-scale discriminator structure is designed,including both global and local branches,to make the network focus on both the overall image and the local image details of the missing areas.At the same time,a Hybrid Loss Optimization(HLO)module is added to constrain the low-level features of the generated image content,promote the generator to improve its ability to generate details,and generate more realistic and clear-textured images.(3)To validate the effectiveness and advancement of the image inpainting method proposed in this paper,qualitative and quantitative comparisons and analyses were conducted between the proposed algorithm and current mainstream algorithms on multiple datasets.Additionally,ablation experiments were performed on the proposed network module.The experimental results show that the RFA-Net image inpainting algorithm proposed in this paper can restore the texture details of damaged areas and generate content consistent with real images,and has a certain degree of advancement in quantitative and qualitative comparison with current mainstream deep learning-based methods. |
