Improved method on image stitching based on optical flow algorithm

Abstract

With the rapid development of technologies based on virtual reality, image stitching is widely used in various fields such as broadcasting, games, education, and architecture. Image stitching is a method for connecting multiple images to produce a high-resolution image and a wide field of view image. It is common for most of the stitching methods to find and match the feature in the image. However, these stitching methods have the disadvantage that they cannot create a perfect 360-degree panoramic image because the depth of the projected area varies depending on the position and direction between adjacent cameras. Therefore, we propose an advanced stitching method to improve the deviation due to the difference in the depth of each area using the pixel value of the input image after the feature-based stitching. After the feature-based stitching method has been performed, the pixel values of overlapping areas in the image are calculated as an optical flow algorithm, then finely distorted, and then corrected to ensure that the image overlaps correctly. Through experiments, it was confirmed that the problem that was deviated from the feature-based stitching was solved. Besides, as a result of performance evaluation, it was proved that the proposed stitching method using an optical flow algorithm is capable of real-time and fast service.

Enhancing Feature Detection and Matching in Low-Pixel-Resolution Hyperspectral Images Using 3D Convolution-Based Siamese Networks

“…These feature matching methods can be divided into two main subsections based on their feature detection method, namely, learning-based feature detection methods and methods which do not incorporate learning. Traditional methods, such as frequency domain methods [ 38 ], spatial domain methods [ 39 ], optical flow methods [ 40 ], Scale-Invariant Feature Transform (SIFT) [ 39 ], Speeded-Up Robust Features (SURF) [ 41 ], Harris corner detection [ 42 ], Features from Accelerated Segment Test (FAST) [ 43 ], Binary Robust Invariant Scalable Keypoints (BRISK) [ 44 ] and Oriented FAST and Rotated BRIEF (ORB) [ 45 ], fall under the heading of non-learning-based methods. With the progress of deep learning technologies, more sophisticated methods of feature detection and matching have been introduced to address drawbacks in non-learning-based methods.…”

Optimizing Image Rectangular Boundaries with Precision: A Genetic Algorithm-Based Approach with Deep Stitching

“…Several other feature detection algorithms have been created subsequently, including speed up robust feature (SURF), oriented BRIEF (ORB), and features from accelerated segment test (FAST), which are used to categorize images based on their distinct properties, such as brightness, image size, and rotation. Once these features have been identified, a descriptor is generated for each feature by calculating its relationship with the surrounding pixels, which describes the feature information [12].…”

Combined 360° Video and Uav Reconstruction of Small-Town Historic City Centers

“…The final projection is considered distortion free, up to a scale factor, and the relative orientation of the different cameras is estimate during the stitching phase (Fangi and Nardinocchi, 2013). Since consumergrade spherical cameras are not designed for metric purposes, several works investigate the possibility of using such instruments for photogrammetric applications, mainly working in two directions: (i) the implementation of more sophisticated algorithms and solutions for image stitching (Lee et al, 2020), and (ii) the improvement of photogrammetric techniques for processing these types of images (Fangi et al, 2018;Janiszewski et al 2022). Indeed, even if the mathematical framework for spherical images is well-defined, some practical aspects may pose significant issues.…”