Lut 250k~400k(FPGA Verification Case)
Flipflop 300k~350k(FPGA Verification Case)
Low-power Real-time 4K Denoising with ASIC
Down to 0.5MB RAM Consumption for 4K Image
BM3D is the upgraded version of the NLM algorithm and has proven to be the best inimage denoising in terms of higher PSNR. Because of its high computationalcomplexity, it's usually used offline. BM3D mainly uses non-local patch matching.
First, it searches for similar patches. While NLM will take weighted average of these similar patches, BM3D transforms these similar patches into a frequency domain, and uses collaborative filtering to reduce noise.
Next it performs an inverse transformation to process those patches.
Point Spread provides deeply-optimized hardware IP Cores of BM3D for RAW images.
Point Spread launched the BM3D IP core for real time 4K RAW images.
It has improved noise reduction by over 30%, and generates less noise in dark scenes.
Point Spread launched the BM3D IP Core for realtime 4K RAW images. It has improved noise reduction by over 30%, and generates less noise in dark scenes.
Original Noisy Image
Output Denoised Image
Single Shot, Sensor Size 1/2'', IMX586 F/1.7, Pixel Size 0.8μm.
Computational Optics Platform
BM3D Denoising IP Core
Computational Imaging Course
Copyright © 2020 - 2022 PointSpread. All Rights Reserved. 点昀技术（南通）有限公司 版权所有