Research On Key Technologies Of Aerial Video Image Mosaic Based On SIFT Features

High resolution aerial images have been widely used in environmental monitoring, natural disaster relief, etc. As for these applications, image mosaic is the fundamental step. Traditional mosaic algorithms, such as local matching and information entropy-based, have been proven to sensitive to rotation, scale and illumination changes. Therefore, in this paper, the Scale Invariant Feature Transform (SIFT) has been used in the process of feature extraction and feature matching, known as the preprocessing stage of image mosaic. Details of our works are given as follows.1. After analyzing the efficiency of SIFT descriptor, a speed up version of SIFT, which aims at video image matching, is presented. The proposed algorithm has a 64 dimension form which is extracted from a sector neighborhood of each keypoint. Moreover, the improved SIFT has been introduced in adjustment model, this model used for determine the affine transformation between images to be matched. The simulation results show that the proposed algorithm speed up the process of image feature extraction, feature matching and image mosaic.2. An improved seam remove algorithm for aerial images has been proposed. Note that multi-temporal image always has different grayscale, different texture. On the other hand, the affine transformation parameters deduced via SIFT features may inaccurate. Are these factors may result in seam in mosaic images. Since aerial images always have distortion and grayscale difference through the seam, traditional algorithms, such as incremental-based strategies will result in multi-seam in the vertical and horizontal directions. Also, dynamic-based methods may fail in this situation. Motivated on this, in this paper, we present a new seam removal algorithm, which aims at transform the complex overlap problem to double overlap. The purpose is to reduce the seams from vertical and horizontal directions to only one. Experimental results obtained from aerial video images show that the proposed method has a better vision appearance, compare to traditional methods.