Research On Local Features Based Image Reconstruction Algorithm

Local features based image reconstruction is an algorithm which uses lo-cal feature information of the original image,together with the large-scale image dataset,to perform accurate image reconstruction so that the reconstructed im-age is similar to the original image and achieves good visual quality. Image reconstruction algorithm has been widely applied in the fields of copy detec-tion,privacy protection,super-resolution reconstruction,3D reconstruction. In this paper, we study reconstruction algorithm framework under the scene of im-age compression application and improve some technical details.The scenario of cloud-base image coding is introduced firstly. The author discusses the complete system process including feature extraction and data compression at the client end, similar image retrieval by using local features on large-scale image dataset and precise image reconstruction from those similar images at the server end. Different modules and corresponding technologies are mentioned. Secondly,we dive into the current architecture of the image reconstruction system from two technical aspects,including (1)the traditional panorama stitching technologies of local features,feature matching,2D trans-form and registration,patch filtering and image fusion.(2)Large scale partial-duplicate image retrieval technologies including bag of visual words model,visual words quantization,locality-sensitive hashing,visual words group and local info based image retrieval.The main work is described as follows:(1)Focusing on the features of large scale corpus, the paper proposes the algorithm of image retrieval based on2D visual words encoding and block based similar query and local matching accu-racy improves.(2)In the part of image stitching, we propose an adaptive thresh-old validation method to filter candidate image patches, the proposed method provides stronger reconstruction evidence and thus improve the performance.(3)Combining these various technologies, we design and implement a complete system including offline training and online reconstruction. Several methods are used to optimize the system for massive data processing, including the use of improved clustering algorithm to generate visual words, fast matching by utilizing the data structure of k-d tree and the use of inverted index for visual words group similarity search.(4)Design and complete multiple sets of com-parative experiments,the results demonstrate that reconstruction has reached a satisfactory level.