The Content-aware Image Scaling Technology Research

With the increasing of diversity and versatility of the display devices which all come in different aspect ratios and resolutions in recent years, content-aware image resizing has become a hot pot in the field of image processing. In this thesis, the key problems of image significance identification, based image significance resizing and resizing effects assessment were profoundly studied. Main works of the thesis are:1. In the stage of the image significance identification, an auto and an interactive recognition method were proposed respectively. The auto method firstly improved the single resolution saliency model based on the rarity of features as multi-resolution model. Secondly, this multi-resolution saliency and image gradient were combined into the image significance estimation. By tested on many images, the auto method results were not only sensitive to image structure information, but also can identify the prominent object as important region. By providing different interactive modes based on pixel and mesh, the interactive method can meet the demands for the significance value expression for image resizing methods based on seam carving and image warping.2. Taking the advantage of the fact that the uniform scaling of keeping source image aspect ratios can preserve the prominent object features, the bi-directional seam carving for image resizing method is proposed. This method divided the image width reduction (enlargement) into three steps:width narrowing (widening), height magnifying (lessening) and uniform scaling of zoom in (zoom out). Do the same case with height scaling. In order to getting the appropriate scaling ratio of the each step, the optimal and interactive bi-directional seam carving methods are presented. The former focuses on research optimization strategy of calculating the seams amount, and the latter emphasizes to study how to real-time response of the user interaction data. The experiment results show that the optimal method can basically obtain the relatively better resizing results, and interactive method can meet the different user subject preference.3. Utilizing the fact that the uniform scaling mesh can keep the image feature in the corresponding region, by controlling the uniform scaling ratio of the image important region, a prominent object size adjustable method for image resizing is proposed. This method presented a measurement model about quad distortion energy by considering both the quad shape and size. Therefore, it can overcome the limitation from the existing mesh warping-based resizing methods: prominent objects will always be reduced or magnified along with the image size reduction or enlargement. Moreover, by adjusting a simple parameter, it also can change the primary object size according to user preference.4. In order to assess the effects and performances of the diversity content-aware image resizing methods, an experimental scheme that combines the linked-paired comparison and randomized block design is presented. This scheme can decrease the comparison pairs number based on the same image which disposed by an experiment participant while keeping the balance among the comparison pairs and experiment participants in the experiment process. Furthermore, this experiment scheme can detect the circular miscalculation in the results from the same participant and the same image in real-time and then show the error results to repeat the test. In the stage of the experimental data analysis, this paper draws the ordination and group conclusion of the 10 image resizing methods by combining diversified classical non-parametric statistical methods.