| Occlusion target recognition plays an extremely important role in the field of artificial intelligence,and the most critical step in solving the problem of occlusion target recognition is image repair.Image repair is a difficult task,and it is a processing technology that can improve image quality in the field of image research.It needs to restore damaged areas in the image based on the existing data of the image.In the actual image acquisition,due to a variety of damaged noise,strong and dark light and physical occlusion and other factors,the quality of the picture has seriously deteriorated.Therefore,the significance of image restoration research is extremely important.Generative Adversarial Networks(GAN)is a neural network architecture proposed by using adversarial thoughts,and is one of the most promising unsupervised learning models in the era of artificial intelligence.In this paper,with the help of GAN's confrontation ideas,we design repair methods for damaged images,and combine the advantages of Squeeze-and-Excitation Networks(SENet)and Self-Attentive(SA)features in feature extraction.Investigating the details of image restoration makes the calculation efficiency and restoration results better during the entire image restoration process.This article mainly completes the following work:First of all,a deep convolution generation based on squeezed excitation block generation image repair model(Occlusion Image Recognition by Deep Convolution Generative Adversarial Networks based on Squeeze-and-Excitation Networks,SE-DCGAN).Redesign the structure of the deep convolution generation adversarial network,and squeeze the excitation block in the third layer convolution layer and the fourth layer convolution layer of the generation model,so that the deep convolution generation anti-network convolution layer extraction Features are more precise.In addition,SE-DCGAN introduces Maxout activation function with powerful fitting function to improve image generation efficiency and avoid image generation redundancy.The simulation experiment results on the three data sets of CelebA,SVHN and anime avatars show that the model successfully predicts a large number of missing areas and achieves coherent image restoration in visually clear content,especially improving the face of the blocked face.Recognition rate.Secondly,a self-attention mechanism and conditional Wasserstein Generation Adversarial Network's Occlusion Image Repair Model(SACWGAN)is proposed.The generator of the conditional Wasserstein generation adversarial network uses theself-attention mechanism to pay more attention to the relationship between the data and the features,and uses the condition label as input to complete the repair work.On the data set CelebA and AR data sets,through the visual display and two important evaluation indicators(PSNR and SSIM),the effectiveness of SACWGAN in processing occlusion tasks was verified.On this basis,combined with SENet's ability to effectively capture key image features in a neural network structure,an occlusion image repair model(SE-SACWGAN)that combines squeezing excitation blocks and self-attention conditions Wasserstein generation confrontation is proposed.The model utilizes the key feature capturing ability of the squeeze excitation block to improve the self-attention condition Wasserstein to generate the occlusion image repair effect against the network.Finally,a prototype system for occlusion image recognition based on SE-SACWGAN is designed.Based on the Django framework,a video image repair demonstration system is designed,and the damaged area in the dynamic video is acquired and repaired by frame interception.This article demonstrates the repair of the damaged video actually shot,and summarizes the effect of repairing the occlusion.The paper has 33 drawings,18 tables and 89 references. |
