There are many open-source low-light AI-ISP models. The table below is a comparison table provided by Copilot.
|
Algorithm |
GitHub |
Type |
i.MX95 NPU Suitability |
FPGA Suitability |
|
MSR (Retinex) |
jsrsinchana/.../MSR-algorithm |
Non-AI (ISP) |
Medium |
Very High |
|
Zero-DCE++ |
arnabroy734/low_light_enhancement |
Lightweight CNN + Curve |
Very High |
Very High |
|
RetinexNet |
weichen582/RetinexNet |
CNN (Retinex) |
Medium |
High |
|
EnlightenGAN |
VITA-Group/EnlightenGAN |
GAN (CNN) |
Very High (lite) |
Low |
|
FLOL |
cidautai/FLOL |
Lightweight CNN |
High |
Low |
|
SNR-aware |
JIA-Lab-research/SNR-Aware |
Transformer + CNN |
Low |
Low |
|
KinD |
zhangyhuaee/KinD |
Retinex + CNN |
Medium |
Medium |
|
RetinexNet-lite |
Derived |
Light CNN |
Medium |
High |
|
EnlightenGAN-lite |
Derived |
Small CNN |
Very High |
Low |
|
Fast LLIE CNN |
Various |
Small CNN |
High |
Medium |
We selected some open-source models and used UVC to perform performance tests on the exip-os08a20 module with no HDR mode. We found that SCI（GitHub - vis-opt-group/SCI: [CVPR 2022] This is the official code for the paper "Toward Fast, Flexib...） computation is relatively small, low-light performance is good in subjective evaluations, and it can basically run on the IMX95. The testing method involves copying the tflite file and test script to the /root/ directory of the IMX95 and running the following command: `python3 test_sci_cvpr_illu_imx95_int8.py --model sci_tpami_illu_imx95_int8.tflite`. The comparison interface shown below is displayed.
