This repository contains the code for a general demosaicing method via joint nonlocal optimization. For more information, please contact Liheng Bian (bian at bit dot edu dot cn).
Figure1. Recovered images of different methods under different MSFAs. The close-ups and corresponding error maps are shown in the right side of each image.
The single-sensor multispectral camera attaches a multispectral filter array (MSFA) on the sensor, enabling to acquire multispectral information in a snapshot. Compared with the other existing multispectral imaging systems that imploy various spectrum splitting and detection equipments, the MSFA based multispectral camera has the advantages of compact size, low weight, low cost,exact registration, strong robustness and full frame rate.
Compared with other methods, the proposed method enables higher-fidelity details and fewer artifacts, and maintains the state-of-the-art performance for different mosaic patterns and channels. Figure1 shows a comparison of the proposed method with other methods.
Please clone this repository by Git or download the zip file firstly.
Run Demo.m file to achieve the demosaicing results of BTES algorithm、DCT algorithm、GAP-TV algorithm and Ours algorithm stored in 'results' file.
Choose datasets of different cases by these two parameters:
`type`: change the arrangement of MSFA, including random MSFA、regular MSFA and btes MSFA;
`num_pic`: change the number of spectral picture for simulation, ranging from 4 to 9.
The func_Ours.m runs the proposed method. It can be used for other tasks as well and only requires inputs of:
`mask`: the arrangement cube of pattern;
`m`: the measurement cube of scene;
`orig`: the image cube of ground truth;
`vinitial`: the image cube of initialization.
The value of input images should range from 0 to 255.
The func_DCT.m is memory consuming for 'toy' dataset with size of 256x256x6. Please ensure at least 32GB memory left.
The func_Ours.m is time consuming for 'toy' dataset with size of 256x256x6. More physical CPU cores conduce to faster reconstruction.
All the experiments are implemented using MATLAB(R) with an Intel i7-9700K processor at 3.6GHz and 64GB RAM. Notice that the func_DCT.m requires large RAM space and the func_Ours.m requires Parallel Computing Toolbox.