基于稀疏编码和矩阵分解的视频去雪算法

Video desnowing algorithm based on sparse coding and matrix decomposition

视频中的雪花会降低视频图像的质量,影响计算机对视频中目标自动检测、跟踪和识别等操作。视频中雪花去除是一项具有挑战性的任务,现存的算法除雪方法效果不佳。根据不同的空间特征,将一个视频中的雪花分为近景雪花和远景雪花。首先,利用低秩矩阵分解提取视频的背景信息。然后,采用多尺度卷积稀疏编码对远景雪花进行检测。利用马尔可夫随机场对运动物体进行建模,之后使用连通域阈值去除被判断为运动物体的雪花。实验结果表明,提出的算法有效地去除了视频中的近景雪花和远景雪花,同时保留了相关的背景和运动物体的信息。

Snowflakes in video reduce the quality of video images which affects the computer's ability to automatically detect,track,and recognize objects in video.Snowflake removal in video is a challenging task,and the existing snow removal methods perform poorly.Snowflakes in a video can be divided into close-up snowflakes and distant snowflakes according to different spatial characteristics.First,we extract background information by employing lowrank matrix decomposition.Then,multiscale convolutional sparse coding is used to detect distant snowflakes.Moving objects are modeled by using Markov Random Field.Additionally,the connected domain threshold in our model is designed to remove snowflakes identified as moving objects.The experimental results show that our proposed model effectively removes the distant snowflakes and the close-up snowflakes in the video,while keeping information of the background and moving objects.