Research On Occlusion Image Inpainting Algorithm Based On Deep Learning

Digital image inpainting technology refers to using the neighborhood information of the image missing area to fill the image missing area according to certain inpainting rules,which makes the observer unable to perceive that the image has been damaged.This technology has been widely used in image deblocking,film and television special effects production,etc.When it comes to image semantic repair or image large-area missing repair,the traditional image inpainting algorithm based on structural texture cannot obtain a satisfactory image missing part repair effect by using the known region information of a single source image.In view of the above problems,this paper has carried out a large number of related research and experiments,the main contents are as follows:(1)Dilated convolution is introduced on the basis of DCGAN network model,which verifies the enhancement of image inpainting by expanding the receptive field of convolution kernel without changing network parameters.(2)According to the down-sampling structure of the generator in the network model,three different convolution kernel sizes are used to extract different horizontal features of the image,and then deconvolution output is performed to generate the image.The influence of the extracted features of different convolution kernel sizes on image inpainting is compared.(3)For the problem of unclear image details after inpainting of the single-channel network structure,discontinuity of the texture of the structure,and the loss of the generator’s loss value during the network training process,Using multi-channel collaborative network structure combined with 1x1 convolution kernel channel-level dimensionality reduction,Parallel training after splicing different network structures,synchronous iteration parameters,and verifying the improvement of the image inpainting details by the multi-channel parallel network structure.(4)Introduce MSE loss based on edge features on the basis of multi-channel cooperative network structure,and combine it with content loss,style loss based on VGG network structure,etc.As the loss function of the generator.Not only did the PSNR value of the inpainting image be 25.412dB on the test set of the Paris street view data set that lacked the central area,but the loss value of the generator during the training process of the network did not show obvious oscillation,and the overall loss declined slowly.(5)For the problem of missing inpainting at any position of the image,two different network structures are proposed,which are trained on the Paris street view dataset and the celebA face dataset.The qualitative and quantitative analysis and comparison of the images after the inpainting of two different network structures found that the network inpainting model based on the external mask data set is better than the network inpainting model based on the movement of the center mask area.In addition,for other types of image missing problems,an experimental comparative analysis was also conducted to verify the generalization of the network model.(6)Based on tkinter developed a set of image inpainting GUI interface,you can add different types of images on the interfaceThe mask is then inpainting,and the inpainting effect is displayed in real time.