Optimization of Video Stability Technology by Integrating Tiny-Res-PWNet Network Model

Abstract

To achieve high-quality video stabilization, this study improves the output structure of the stabilization network and adds Fourier spectrum constraints and local motion constraints, while optimizing the affine matrix parameters. In addition, this study replaces the encoder convolutional layer with a residual module and combines feature fusion to process feature information extracted from different network layers. This can achieve optimization and lightweight processing of pixel by pixel stable network models. The results show that compared to the previous pixel by pixel stable network model, the stability evaluation index of the improved pixel by pixel stable network model has increased by about 3.7%. Compared to the pixel by pixel stable network model encoder, the parameter count of the pixel by pixel stable network model encoder fused with residual module is reduced by 12.1%, the model size is reduced by 11.7%, the floating-point operation is reduced by 13.2%, and the running frame rate is increased by 5.6%. The lightweight pixel by pixel stable network model can achieve a frame rate of 131.2 at high-performance operation, far higher than the 83.1 of the pixel by pixel stable network model. The outcomes showcase that the network model is an effective optimization method for video stabilization technology and can be applied to many real-time video processing scenarios. This helps to improve the technical level and application effectiveness in this field.