ldr2hdr, fldr2hdr, optical filter, U-net
DITMnet：ldr2hdr via CNNs, detail in paper 《Method for Reconstructing a High Dynamic Range Image Based on a Single-shot Filtered Low Dynamic Range Image》accepted in Optics Express
original HDR dataset come from online: Fairchild-HDR {http://rit-mcsl.org/fairchild//HDRPS/HDRthumbs.html}, Funt-HDR {https://www2.cs.sfu.ca/~colour/data/} and DML-HDR {http://dml.ece.ubc.ca/data/DML-HDR/}.
2020-01-09
We have made further adjustments and optimizations based on the previous manuscripts and code. The details are as follows:
-
Added virtual filter camera code for simulating physical optical filters, see "generate_data.py" or "generate_data_v1.py"
-
Further optimized the code of the network structure to make it more clear and concise, see "src/archs.py"
-
Integrated training, testing, and prediction code, see "main.py"
pipeline:
step 0: generate training pairs by "generate_data.py" or "generate_data_v1.py"
step 1: create network, in this paper, we create a multi-branch and multi-ouput CNNs network, detail in "src/archs.py"
step 2: train this network by "main.py". model type=0 Note that in this paper, we use two datasets, the first had been training and the last dataset (Funt-HDR) had no training.
step 3: test this network by "main.py". model type=1
step 4: predict the results by "main.py". model type=2
step 5: performance comparison by Matlab code