Research On Semantic Image Inpainting Based On Object Optimization

As a very important research topic in the field of digital image processing,image inpainting technology has been rapidly developed and applied in recent years.Image inpainting refers to the process of using computer technology to repair the image degradation caused by the process of acquisition,storage and transmission,which possesses great significance in scientific research,medical and other fields.Traditional inpainting algorithms mainly accomplish inpainting by diffusing and matching low-order image information such as color,texture and structure.However,these algorithms suffer from more limitations and have bad performance.In recent years,with the development of large data processing technology and the growing maturity of hardware resources.Deep learning surpasses traditional image processing algorithms in the area of image processing like image classification,object detection,image segmentation and image inpainting with absolute advantages.In the field of image inpainting,many excellent models have been proposed by combining encoder-decoder pipeline with Generative Adversarial Network.These models have made a qualitative leap in the effect and ability of image inpainting compared with the traditional algorithms mentioned above.However,there are some unhandled problems remained in the image inpainting task When the missing region is set to be a part of a whole individual with a small proportion in an image,the inpainting model will lead to a very absurd result.The essence of this problem is that the generalization ability of data sets used in training is too strong,which makes the algorithm unable to calculate for every individual in the image.To solve this problem,this paper proposes a semantic image inpainting method based on objective optimization,which can improve the semantic ambiguity of the inpainting results to a certain extentThe algorithm proposed in this paper mainly includes two aspects:one is to establish an improved generative inpainting model,which can capture more semantic information by adding parallel void convolution and style loss;the other is to construct a semantic inpainting method based on object optimization by combining the inpainting model with the semantic segmentation model.What's more,this paper also provides an end-to-end one-key image inpainting platform integrating many mature inpainting algorithms.The backend integrates a variety of mainstream deep learning inpainting algorithms,which can make custom mask during the process of inpainting,and shield the differences in the details of various algorithms.The experimental results show that the proposed algorithm has higher availability and fault tolerance than the existing mature algorithms in semantic inpainting and object removal tasks.In addition,the effects of the size of the missing area and the step size of the dilation convolution on the inpainting result of the model are also discussed through comparative experiments.