Sparse Coding And Frequency Domain Directional Filtering Based Image Rain And Snow Removal

Adverse weather conditions, such as rain or snow, severely degrade computer vision systems' performance. Vision systems cannot offer reliable results of computer vision algorithms such as target detection, object recognition and tracing and feature extraction. As a result, it is necessary to take the rain removal as a preprocessing step to improve the performance of computer vision systems.We first analyze and compare the existing rain and snow removal algorithms which can be classified into four categories, spatial-temporal domain based, chromatic property based, frequency domain based and matrix factorization based. Then we propose some improvements on existing sparse coding based single image rain and snow removal algorithm. First we use guiding filter rather than bilateral filter as preprocessing step to separate the image into high-frequency part and low-frequency part. Then we reclassify the misclassified atoms with no rain or snow based on the proportion of edge points in an atom. Besides that, we use color mask to further improve rain and snow removal performance.Moreover, we propose a new frequency domain directional filter based rain and snow removal algorithm. To be specific, we first convert the input color image into YCbCr color space and deal with the Y-component image only. Then we compute the dominant direction of the global edge based on the local Histogram of Oriented Edge to determine the direction of the rain and snow streaks, which is perpendicular to that of wedge-style spectrum of rain and snow in frequency domain. Design the corresponding 2D frequency domain directional filter using McClellan transform, and then utilize the filter to remove the frequency spectrum of rain and snow from that of Y-component. Finally, we reconstruct the output image by combining the filtered Y-component with original Cb and Cr component. The experimental results show that our algorithm maintains more details while works faster than other existing algorithms.