Dimensionality reduction algorithms ha ve been applied widely in computer vision,medical image processing,video process ing,face recognition,image retrieval etc.However,some problems need to be fixed. First of all,how to get the domain size and intrinsic dimension is first primary problem to be fixed,which is seriously restricted the rapid development.Tradito nal method used the K-Nearest Neighbor to search the neighborhoods of each imag e sample.But it costs much time sometimes.In this paper,we aim at the problem of finding intrinsic structure and domain size automatically for high dimensional image data.We present a new technique which can get intrinsic dimension and neig hborhood size automatically and adaptively.The algorithm can be used for extract ing local features from images.Firstly,we made linear reconstruction based on th e nearest neighbor distance for image feature extraction,and optimize the distri bution on whole manifold,then we get expression function in which variable is th e optimal linear reconstruction for locally low dimensional feature.Lastly,minim iing the variance of the function is to get aotomatic selection strategy.Experim ents show that the algorithm is not only simple but also high matching rate and low computational complexity.