Context-Aware Hierarchical Transformer For Fine-Grained Video-Text Retrieval
Mingliang Chen, Weimin Zhang, Yurui Ren, Ge Li
-
SPS
IEEE Members: $11.00
Non-members: $15.00Length: 00:08:30
Existing snow removal approaches ignore specific characteristics of the snowflake itself, leading to insufficient snowflake feature representation and further incomplete snow removal. We propose a novel snow removal algorithm considering the diversity and complexity of snow, named as DCSNet. For diversity, we construct an adaptive fusion feature pyramid structure, characterizing the four elements of snowflakes(i.e., shape, size, transparency and direction), guiding the following snow removal process with more accurate direction location and salient shape features. We further capture and fuse cross-scale interaction information from the four elements of represented snowflakes more comprehensively. For complexity, we design a progressive recovery module to decompose snowflakes layers stage by stage, allowing the previous feature maps to interact with degraded images for achieving clearer snowflake removal. Extensive experimental results show that DCSNet outperforms the state-of-the-art desnowing algorithms by 3 dB increase in PSNR on three representative datasets. The source code is available at https://github.com/Xjg-0216/DCSNet.