Research On Electronic Stabilization Technology Based On Lightweight Convolutional Neural Networks

During the process of camera shooting,external vibrations can cause the video image to shake,which not only affects the viewing effect but also reduces the accuracy of subsequent image processing.Rapid and effective stabilization technology is of great research significance and practical value.Excellent video stabilization technology should meet the requirements of stability,real-time performance,and robustness.This thesis focuses on electronic stabilization technology,including traditional electronic stabilization algorithms and deep learning stabilization algorithms.This thesis conducts in-depth research on PWStable Net,a deep learning stabilization network that meets the above three requirements,and proposes optimization improvements based on application scenarios.The main contents of the work are as follows:In response to the issue of local distortion in the output videos of PWStable Net,a stable image network,a proposed improvement scheme was developed after a detailed analysis of the causes of distortion.To address the challenge of real-time stabilization processing on low-performance hardware devices,a lightweight improvement scheme was proposed,focusing on compressing the network structure and enhancing processing speed.The specific optimization improvements proposed in this thesis are as follows:(1)Frequency domain constraints and local motion constraints have been added to the neural network as a loss function,guiding the network to converge towards local stability and suppressing local distortion in the images.(2)The model used for coarse motion estimation in the neural network has been modified,replacing the 6-parameter affine transformation with a 4-parameter similarity transformation.This change enables the network to learn image motion patterns explicitly rather than implicitly,suppressing irregular motion.(3)The bottleneck residual module has been introduced to replace the downsampling convolutional layers,optimizing the encoder structure of the stabilization network.This prevents network degradation and accelerates the processing speed.The optimized structure is referred to as Res-PWNet.(4)Based on Res-PWNet,some standard convolutions have been replaced with depth-wise separable convolutions.This utilizes the characteristics of depth-wise separable convolutions to significantly reduce the model’s parameter and computational complexity.The optimized structure is named Tiny-Res-PWNet.Experimental tests have shown that the proposed improvement method greatly reduces local distortion in stabilized videos,leading to improved subjective visual quality and enhanced objective performance indicators.The proposed method outperforms traditional stabilization methods in stabilizing video data under lowquality shooting conditions,demonstrating better robustness and meeting the requirements of video stabilization.Compared to the previous version,the Tiny-ResPWNet stabilization network model proposed in this thesis achieves significantly improved processing speed,reaching a real-time stabilization requirement of 32.1 FPS on medium-performance devices.In typical applications such as airborne remote sensing video stabilization,the stability and robustness of the Tiny-Res-PWNet stabilization algorithm also meet the desired requirements.