A Posteriori Outlier Rejection Approach Owing to the Well-ordering Property of a Sample Consensus Method for the Stitching of Drone-based Thermal Aerial Images

Abstract

In this study, the authors generate panoramic images using feature-based registration for drone-based aerial thermal images. In the case of drone aerial images, the distortion of the photographing angle due to the unstableness in the shooting altitude deteriorates the performance of the stitching. Furthermore, for the thermal aerial images, the same objects photographed at the same time zone may have different colors due to the relative temperature, which may lead to a more severe condition to be stitched. Applying the scale-invariant feature transform descriptor, they propose a posteriori outlier rejection scheme to estimate the hypothesis of the mapping function for the stitching of consecutive thermal aerial images. By extension of the method of optimal choice of initial candidate inliers (OCICI) and a posteriori outlier rejection scheme using cross-correlation calculus, the authors obtained elaborate stitching of thermal aerial images. Their proposed method is numerically verified for its quality by comparing it with other possible approaches of post-outlier rejection treatments employed of OCICI. Also, after the Poisson blending using the finite difference method is conducted, the stitching performance is compared with some benchmark software such as Matlab-toolbox, OpenCV, Autopano Giga, Hugin, and PTGui.

A Cognitive Sample Consensus Method for the Stitching of Drone-Based Aerial Images Supported by a Generative Adversarial Network for False Positive Reduction

“…The general approach for stitching images involves identifying key points and comparing their local descriptors to determine the transform of the key points’ geometric correspondence to generate the resampling rule of the sensed images [ 1 , 2 ]. In our study, we use SIFT as the descriptor, which is applied in many studies for its well-known geometric invariance properties [ 26 , 28 ]. Figure 1 shows the reference image and sensed image with the marks of the SIFT points, where the images are resized into a 200 250 resolution; we present 50 random SIFT points and the lattices of the orientations.…”

“…The descriptors in both images are then matched to form a set of correspondences that estimates the parameters of the resampling rigid transform (an affine map made of a 3 3 matrix [ 4 , 26 ]) using RANSAC.…”

“…For example, we apply deep learning to verify RANSAC, a method widely used to find the true inlier of the mapping function, and we improve its performance by linking it with other a posteriori outlier rejection methods, such as optimal choice of initial correspondence inliers (OCICI) [ 4 , 26 ]. In the process of determining the hypothesis found through RANSAC, its authenticity is determined by training a generative adversarial network (GAN) [ 27 ].…”

“…To implement the proposed method for experimental verification, we apply it to drone-based aerial images. Previous studies have shown that it is relatively difficult to solve the correspondence problem for aerial images based on drones [ 4 , 26 ], and our proposed method shows a way to improve the efficiency of the existing methodologies supported by deep learning technology. Actually, our proposed deep learning methodology deals with the topography of the estimated resampling operators, such as affine maps and not the images; hence, from the method’s inherent properties, it may be applicable for correspondence problems with various types of images.…”