Research On Retrieval Based Multi-Option Image Completion

Digital image completion is one of image processing technique to make use of known region to complete the missing region. It requests the completed image maintain good visual condition, minimize the artificial details after the process of completion. Digital image completion model could be classified into self-content based completion model and material image based multi-option completion model. Compared to self-content based completion model, there are obvious advantages by using material image based multi-option completion model.On the one hand, it’s effective to solve the problem of large-area missing. On the other hand, it offers more completion options to uncompleted image, multi-option image completion results are more diversity and free. In the media age, many image processing techniques play an important role. As one important technique in the field of image processing, material image based multi-option completion model provides important technical support to media applications. Its basic process could be summarized as three sections in the following:Firstly, we need to retrieve material images (they are as similar as uncompleted image and can be used as completion material) from massive image database, secondly, we need to extract the material regions(the specific regions used for completion in the material images) from material images; thirdly, we need to use material region to complete the missing region. However, how to retrieve enough and accurate material images, how to extract the materal regions exactly, how to reduce the artificial details (the gradient of edge is changed obviously, there are obvious color, texture, noise differences between foreground and background) are key problems in the process of completion. Traditional methods often can’t solve these problems effectively and reduce the quality of completion. In this paper, we study these questions in depth and do some further research on "retrieval based multi-option image completion", the main contributions are in the following:(1) For the problem the accuracy of material image retrieval is not realistic, we proposed one optimum material image selected model. Firstly, we use different category joint distribution probability to do preliminary retrieval. In the process of it, we make use of label information to delete some unrelated images and reduce some useless work effectively. Secondly, we use K-means algorithm to classify the images in the database. The images are classified into scene images and object images. Based on different categories, we use different feature descriptors to handle them. Lastly, we use improved spatial pyramid matching function to achieve spatial matching between uncompleted image and candidate material images exactly. The three sections are related closely, it solves the problem that the accuracy of material image retrieval is not realistic. At the same time, it contributes to offer enough and accurate material images for completion effectively.(2) For the problem it’s difficult to extract the material region exactly from retrieval image, we proposed one method that can extract the material region exactly. Firstly, we use multi-scale detail preserved and multi-layer smoothing method to optimize the image, which can make the foreground and background keep enough differences. At the same time, the material region’s color and texture are consistent. After optimization, the material region can be extracted easily. For the high-quality matting method(in some special cases, material region need high precision extraction), we make use of optimum learning samples to achieve it. It is effective to solve the problem that we can’t extract the target region exactly and offer enough high quality regions to completion technique.(3) For the problem completion process is complicated and some artificial details exist after completion, we proposed one high-quality completion model. Firstly, we make use of improved foe(field of experts) method to complete the missing region, it is effective to complete it. In the process, optimum learning instance selected model is used to reduce the work-load. However, simple completion based on foe method is lack of diversity and creativity. We use extracted region above to achieve multi-option completion.In the process of completion, we make use of multi-scale color matching, multi-scale texture matching, multi-scale noise handling techniques. These techniques can harmonize the material region and original image effectively so that it contributes to keep the good visual condition. The several sections above are related closely and it’s effective to optimize the completion results.