A Super Pixel Segmentation Algorithm Based On Hadoop

With the diversified development of the type of information,the comfort degree of the human-machine interaction is effected by the quality of image which is the direct expression for transmitting and storing the information.As the prior step of the image processing,the image segmentation plays an important part,with the upgrading of the hardware facility,more and more information included in the images,the focus of the current research in segmentation algorithm is how to design a well adaptive algorithm to the image of large size and high resolution.The Hadoop is wildly used as the big data processing platform at present,whose prac-ticality and efficiency is recognized by the majority of researchers.But because of the original purpose of Hadoop is to process text data,it has only a few data types.From the problems up above,the author presents a super pixel segmentation on Hadoop based on a new I/O on the image.In this paper,the author used super pixels instead of the original pixels as the segmentation processing elements and combine the characteristics of Hadoop and the super pixels.The optimization of this paper from three aspect which are the initial splitting algorithm,the super pixel segmentation in the task nodes and the region merging of the super pixel area.In the first aspect,the author proposed a dynamic adaptive algorithm based on multi task which can reduce the coupling of the blocks in HDFS(Hadoop Distributed File System)and task arranging.In the second aspect,the author designed a distributed watershed algorithm to produce the super pixels using the constraint of the gradient and distance in Map nodes.Finally,a distributed merging strategy was be proved and deployed in the Reduce nodes.In the experiment all the small size images are from the Berkeley image database,the large size images are from the satellite named QuickBird..The final experiment shows that the image segmentation effect of the algorithm to different sizes are satisfactory,and the algorithm has obvious advantages in the time index.