圖像去雨對於自動駕駛系統非常重要，因為這些系統需要清晰的圖像來計算駕駛狀況，例如先進駕駛輔助系統(ADAS)。近期，已經有許多關於圖像去雨的研究工作。但是這些研究都沒有關注影像去雨系統的推理速度。此外，其中大多數的研究都使用合成的下雨圖像來訓練模型，這會造成對真實下雨圖像的泛化性較低。為了解決模型的推理速度，我們提出了輕量級通道注意模塊，並將火模塊引入我們的模型之中，來減少模型中的參數數量。此外，基於相同目的，我們也刪除了一些卷積核過濾器。最後，儘管我們的模型生成的圖像品質較低，但我們部屬在NVIDIA AGX Xavier開發者套件上的模型比基準模型快2.74倍。為了解決模型泛化問題，我們在訓練集中混合了真實和合成的下雨圖像。當我們的模型部署在 NVIDIA 的下一代嵌入式設備 Jetson AGX Orin 開發者套件上時，我們希望我們的模型能夠即時處理去雨（即超過10 FPS）。;Image deraining is essential for autonomous driving systems which may require clear images to calculate driving situations, such as ADAS. Recently, many research works have been proposed for image deraining. However, none of them focuses on the inference speed. In addition, most of them use synthetic raining images to train their models, which leads to lower generalization for real raining images. To address the inference speed, we propose Lightweight Channel Attention Block and introduce fire modules into our model to reduce the number of parameters in the model. In addition, several kernel filters are removed from subnetworks for the same purpose. At the end, although our model generates deraining images with slightly lower quality, our model deployed on NVIDIA AGX Xavier Developer Kit is 2.74 times faster than the baseline model. To address the model generalization issue, we mix the real and synthetic raining images in the training set. We expect that the our model can deal with deraining in real-time (i.e., exceeding 10 FPS) when our model is deployed on the next generation embedded device of NVIDIA, Jetson AGX Orin Developer Kit.
