Deep Convolutional Neural Network for Image Deconvolution
Authors: Li Xu, Jimmy SJ Ren, Ce Liu, Jiaya Jia
NeurIPS 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our system outperforms previous ones especially when the blurred input images are partially saturated. We have presented several deconvolution results. Here we show quantitative evaluation of our method against state-of-the-art approaches |
| Researcher Affiliation | Collaboration | Li Xu Lenovo Research & Technology xulihk@lenovo.com Jimmy SJ. Ren Lenovo Research & Technology jimmy.sj.ren@gmail.com Ce Liu Microsoft Research celiu@microsoft.com Jiaya Jia The Chinese University of Hong Kong leojia@cse.cuhk.edu.hk |
| Pseudocode | No | The paper describes methods in prose and diagrams but does not include any explicit pseudocode blocks or algorithms. |
| Open Source Code | Yes | Project webpage: http://www.lxu.me/projects/dcnn/. The implementation, as well as the dataset, is available at the project webpage. |
| Open Datasets | Yes | We use 2,500 natural images downloaded from Flickr. Two million patches are randomly sampled from them. The implementation, as well as the dataset, is available at the project webpage. |
| Dataset Splits | No | The paper discusses training and testing but does not explicitly specify a validation set or its split proportions/counts for reproduction. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments (e.g., GPU/CPU models, memory). |
| Software Dependencies | No | The paper mentions general techniques and existing neural network architectures but does not specify software dependencies like library names with version numbers. |
| Experiment Setup | Yes | The first hidden layer h1 is generated by applying 38 large-scale one-dimensional kernels of size 121 × 1, according to the analysis in Section 5.1. The values 38 and 121 are empirically determined, which can be altered for different inputs. Fine tuning is performed by feeding one hundred thousand 184 × 184 patches into the whole network. The deconvolution CNN is trained using the initialization from separable inversion as described before. |