A Lightweight Channel-spatial Attention Network for Real-time Image De-raining
Yirui Zeng, Zhengming Ma · 2019
Image de-raining aims to eliminate rain streaks captured by outdoor equipment such as video surveillance, remote sensor and automatic pilot. Recently, a de-raining method called non-locally enhanced encoder-decoder network (NLEDN) has achieved reliability performance. Nevertheless, it is very time consuming (2.2571s per image) and takes up memory so that it cannot be applied to mobile devices to process image in real-time. To solve this problem, we design a lightweight channel-spatial attention network that is 55 times faster (41ms per image) and memory saving. The most advanced performances are achieved in most de-raining data sets. More specifically, we design a channel-spatial attention dense block (CSADB). The channel attention operation will be carried out together with the spatial attention. Our experiments demonstrate that the network can learn more effective features by this way. In order to make our proposed method more lightweight, the depthwise convolutions are adapted in each block to reduce parameters. We conduct experiments on four public synthetic datasets to demonstrate the effectiveness of our proposed method, which achieve excellent performance. And the real-world de-raining results are also tacked into comparison. Moreover, an additional experiment demonstrates that our method also works well on face hallucination task. The relevant code and trained models will be available in GitHub soon.