Research On Invariant Feature Extraction Of Complex Target Based On Multiscale Analysis

The targets human faced are becoming more and more complicated in human daily in the modern society.But there exists some invariant features no matter how complicated the targets become.When the invariant features are extracted in the complicated targets,the features to be detected are different on different scales.Thus it is necessary to detect the features on different scales in order to make sure that the features we detected can describe the targets accurately and effectively.The main contents and the innovation of this paper are as follows:(1)The traditional feature extraction methods can not extract the feature points on different scales.Aiming at the problem,it is proposed that the multiscale analysis method is combined with the feature extraction methods in this paper.So the feature points can be detected on different scales.(2)The morphological gradient operator does not accumulate the noise in the target images.So the contour features of the face images is extracted through the morphological gradient operator on the coarse scale.And a new face database called the ORL contour face database is formed.Then the SIFT algorithm and the CNN model are used to identify and match the contour face database respectively.The recognition efficiency is improved by 1.78% and the the recognition speed is improved by 34.35% when using the SIFT algorithm in the ORL contour face database.And the recognition efficiency is improved by 2.33% and the the recognition speed is improved by 34.96% when using the CNN model in the ORL contour face database.(3)The texture information of the face target image is largely lost when using the contour face database so that the identity errors may happen.Aiming at the problem,the face images which has large similarity with the detected target image is extracted from the ORL face database through similarity and comparison after extracting the contour of the ORL face database.And a new face database called the ORL reduced face database is formed.Then the SIFT algorithm and the CNN model are used to identify and match the reduced face database respectively.he recognition efficiency is improved by 4.78% and the the recognition speed is improved by 45.80% when using the SIFT algorithm in the ORL contour face database.And the recognition efficiency is improved by 7.20% and the the recognition speed is improved by 46.77% when using the CNN model in the ORL contour face database.