基于文本特征融合的双流生成对抗修复网络OA

In order to solve the problems of insufficient feature mining and incomplete semantic expression in deep learning technology,and to eliminate the phenomenon of artifacts or fuzzy texture in repaired images,a dual-stream generative adversarial restoration network with context feature fusion is constructed in the present paper,which is used to realize reconstruction,perception and style loss compensation,so as to achieve global consistency of repaired images.Integrated with the attention mechanism,the U-Net is used as the backbone network to fully extract the image structure and texture features in this network.The upper and lower text feature fusion network is used to fully excavate the high-level semantic and contextual relationship of the feature information,to realize the filling and fine restoration of the structure and texture features of the empty region.The structure and texture dual-stream discriminator is used to estimate the features of texture and structure and distinguish the real image from the generated image.A semantic-based joint loss function is used to enhance the semantic authenticity of repaired images.The CTSDG algorithm with the best performance among the comparison algorithms is compared based on CelebA and Places2 datasets.The values of PSNR and SSIM on CelebA are increased by 2.74 dB and 5.80%respectively,while the value of FID is decreased by 4.02.Similarly,the values of PSNR and SSIM on Place2 increases by 4.15 dB and 3.33%respectively,while the value of FID decreases of 2.33.Therefore,the improved image restoration method shows better objective evaluation indexes,which can repair the structure and texture information of damaged images more effectively,and makes the performance of image restoration better.