基于局部运动约束的改进PWStableNet电子稳像网络

Improved PWStableNet Electronic Image Stabilization Network Based on Local Motion Constraints

为了改善PWStableNet稳像神经网络输出的视频图像局部易发生失真的缺点,提出了2种新的约束方式作为改进,以损失函数的形式来抑制输出视频中的过量运动,阻止局部失真的发生。主要是2种约束方案的设计:一是对翘曲场进行频域分析,鼓励低频信息抑制高频信息;二是对生成的稳定帧进行分析,计算局部运动的剧烈程度,鼓励平稳运动抑制剧烈运动。经过实验数据对比,在主观效果上,改进后网络输出的帧间差分图更接近于真稳定视频,过量的局部运动得到抑制;在客观评价指标上,改进后的网络多数性能指标高于改进前,性能有效程度提升平均值为7.73%,说明改进后神经网络的综合质量优于原网络。

In order to improve the local distortion of the video image output by PWStableNet image stabilization neural network,two new constraint methods are proposed as improvements to suppress excessive motion in the output video by the form of loss function and prevent the occurrence of local distortion.The main work is the design of two constraint schemes.One scheme analyzes the warpage field in frequency domain to encourage low-frequency information and suppress high-frequency information;the other analyzes the generated stable frame,calculates the intensity of local motion,and also encourages stable motion and suppresses violent motion.Through the comparison of experimental data,in terms of subjective effect,the inter-frame difference diagram output by the improved network is closer to the true stable video,and the excessive local motion is suppressed;in terms of objective evaluation indicators,most performance indicators of the improved network are higher than those before the improvement,and the average improvement of performance effectiveness is 7.73%,which shows that the comprehensive quality of the improved neural network is better than that of the original network.