An Intelligent Hybrid Machine Learning With Meta‐Heuristic Optimization Algorithms for Enhancing Real‐Time Video Stabilization

ABSTRACT In the modern era, video stabilization is one of the essential advancement features of digital video processing equipped with 5G technology. Also, this technology leverages the intelligent software innovations to deliver high quality and smooth video recording experiences. Despite advancement in machine learning (ML) algorithms for video stabilization, there are numerous challenges, especially when applying 5G technologies like stable and unstable videos for training performance. Consequently, video stabilization includes complex analyses such as frame interpolation and motion assessment. Moreover, the advanced stabilization modes are developed to analyze the motion data. Nevertheless, they decrease or fail to calculate the features and provide poor results. To overcome these issues, an adaptive video stabilization methodology is proposed. In the proposed method, a novel Convolution Neural with StabNet based Hawks Optimization (CNSbHO) algorithm is introduced. In this research, hand‐held video clips generally suffer from unwanted video jitters due to unbalanced camera motion. Therefore, 5G ultra‐low latency with respect to drone footage video feeds is taken as the stabilization process. Then, a pre‐processing Gaussian filter was enabled to enhance consistency and quality. Hereafter, a Convolution Neural Network (CNN) was used to extract the features, and motion estimation is also done in this section with feature tracking point of CNN. Furthermore, end‐to‐end stabilization strategy as StabNet model can provide stabilized frame outputs. Then, the Harris Hawks Optimization (HHO) algorithm was used to enhance the accuracy of the entire performance. The developed CNSbHO strategy was implemented in Python and validated using the 5G traffic datasets. In order to validate the effectiveness of the developed strategy, we selected the traditional algorithms for the comparison in terms of learned perceptual image patch similarity (LPIPS), structural similarity index (SSIM), accuracy, and peak signal‐to‐noise ratio (PSNR). The comparative assessment confirms that the proposed method outperforms conventional stabilization techniques, making it a reliable solution for real‐time video processing tasks.