Research On Progressive And Parallel Retrieval Algorithm For High Quality Images

Thanks to the diversity and popularity of image acquisition technique,data driven methods for image analysis and edit become popular.However,the explosive growth of images also presents challenges.To help users retrieve expected images quickly and effectively is one of the most difficult.The retrieve cue is the first problem when designing an image retrieval method.It is difficult for a single cue to express the user's intention completely,thus combining multiple cues becomes necessary.Another problem is that the content-based matching process takes a long time.In the case of massive image data,multi cue model is undoubtedly a great challenge for real-time interaction.On the other hand,the purpose of image retrieval is usually to provide material for further image analysis,editing and digital content creation.However,most of the current retrieval methods cannot guarantee the quality of the image which is usually required for many image analysis and editing tasks.In this dissertation,we propose a progressive retrieval method for high quality images,targeting on the problems that the existed retrieval methods cannot guarantee the image quality and the time consuming of the feature matching for the massive material.Images are progressively filtered out according to a series of criteria:we firstly pick up candidates from the huge image database according to a quick textual comparison with images' tags;then we calculate the high saliency objects of each image,and filter out those objects with low saliency scores;for each candidate image,we evaluate an aesthetic score from the aesthetic standards of the three points,diagonal rule and visual balance,and filter out the part that does not meet the threshold of aesthetic standards;next,we calculate the boundary clarity of each salient object and filter out the object with low clarity;for the remaining objects,we expand the surrounding area and count the number of segment in this area by a graph cut method;The greater the number,the higher the complexity,and then filtering out the image that is too complex;for each of the remaining images,we firstly expand the high saliency region by morphological dilation and then cut the scene item according this expanded region.Finally,we measure the consistency between the user drawn contour and the scene item contour,and return the result set that best matches the user's intent.The whole algorithm involves a lot of calculation and matching steps based on the image feature,so performance will become a key issue.We parallelize the whole framework in a MapReduce framework to further improve the performance.We validate the performance,accuracy of the algorithm and quality of retrieval result in various experiments.We also demonstrate the potential of our technique with an image synthesis prototype system.