Hybrid CNN-Transformer Feature Fusion for Single Image Deraining

Authors: Xiang Chen, Jinshan Pan, Jiyang Lu, Zhentao Fan, Hao Li

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive evaluations show the effectiveness and extensibility of our developed HCT-FFN. The source code is available at https://github.com/cschenxiang/HCT-FFN. ... Experiments Experimental Settings Datasets. We conduct deraining experiments on four public rain streak datasets, including Rain100L (Yang et al. 2017), Rain100H (Yang et al. 2017), Rain12 (Li et al. 2016), and Rain DS-Real (Quan et al. 2021). ... Implementation details. The proposed network is implemented in Py Torch framework using Adam optimizer with a learning rate of 0.0001 to minimize Ltotal by 400 epochs.
Researcher Affiliation Academia Xiang Chen1, Jinshan Pan1*, Jiyang Lu2, Zhentao Fan2, Hao Li1 1 School of Computer Science and Engineering, Nanjing University of Science and Technology 2 College of Electronic Information Engineering, Shenyang Aerospace University {chenxiang, haoli}@njust.edu.cn, sdluran@gmail.com, {lujiyang1, fanzhentao}@stu.sau.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at https://github.com/cschenxiang/HCT-FFN.
Open Datasets Yes Datasets. We conduct deraining experiments on four public rain streak datasets, including Rain100L (Yang et al. 2017), Rain100H (Yang et al. 2017), Rain12 (Li et al. 2016), and Rain DS-Real (Quan et al. 2021). ... Rain100L and Rain100H contain 1,800 image pairs for training and 100 image pairs for testing. Rain12 contains 12 light rainy images.
Dataset Splits Yes Rain100L and Rain100H contain 1,800 image pairs for training and 100 image pairs for testing. Rain12 contains 12 light rainy images. ... The loss trade-off parameter is defined via cross validation using the validation set, and the whole pipeline is performed in an end-to-end fashion without costly large-scale pretraining (Chen et al. 2021).
Hardware Specification Yes During training, we run all of our experiments with batch size of 4 and patch size of 128 on one NVIDIA Tesla V100 GPU (32G).
Software Dependencies No The paper mentions 'Py Torch framework' but does not specify a version number or any other software dependencies with version numbers.
Experiment Setup Yes The proposed network is implemented in Py Torch framework using Adam optimizer with a learning rate of 0.0001 to minimize Ltotal by 400 epochs. In our model, {N1, N2, N3} are set to {4, 3, 4}. During training, we run all of our experiments with batch size of 4 and patch size of 128 on one NVIDIA Tesla V100 GPU (32G). In the Da Mo E module, we set k = 8 for the number of experts and T = 32 for the weight matrixs. Each convolutional layer has a C = 16 filter with stride of 1. In the Ba Vi T module, we set k = 3 for the kernel size and s = 4 for splitting segment. The number of heads in MHSA is set to 8. For data augmentation, vertical and horizontal flips are randomly applied. The loss trade-off parameter is defined via cross validation using the validation set, and the whole pipeline is performed in an end-to-end fashion without costly large-scale pretraining (Chen et al. 2021).