introduction

a simple binary neural network for SISR(base on EDSR official code)

if you want to try different binary function, you need to :

1. add new script to model/binarize and modify model/init.py to choose your binary function.
2. edit run.sh(view this script for more detail), choose one function you want to run.
3. run sh run.sh

performance

an example binary function: bianrize/WApproxASTE.py:

weight binarization

forward:

backward,use approxsign (bireal-net's method)

activation binarization

forward:

backward,STE

network modification

add an Tanh() function before binact. It helps to transfer a nonlinear gradient after STE.

srresnet

Method	Bit	Set5	Set14	B100	Urban100
	W/A	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
srresnetx2	32/32	37.889/0.958	33.4/0.915	32.077/0.896	31.602/0.922
Oursx2	1/1	36.345/0.934	32.221/0.876	31.364/0.877	29.407/0.883
Oursx2(DBSR)	1/1	37.502	33.085	31.82	30.881
FPx4	32/32	32.066/0.890	28.497/0.778	27.516/0.731	25.858/0.778
Oursx4	1/1	31.232/0.848	28.047/0.729	27.215/0.699	25.075/ 0.727
Oursx4(DBSR)	1/1	31.666	28.282	27.333	25.419

compare to Efficient Super Resolution Using Binarized Neural Network(Ma et.al.)

Method	Bit	Set5	Set14	Urban100
	W/A	PSNR	PSNR	PSNR
Ma et.al. x2	1/1	35.66	31.56	28.76
Oursx2	1/1	36.345	32.221	29.407
Ma et.al. x4	1/1	30.34	27.16	24.48
Oursx4	1/1	31.232	28.047	25.075

edsr

Method	Bit	Set5	Set14	B100	Urban100
	W/A	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
edsrx2	32/32	37.931/0.958	33.459/0.915	32.102/0.896	31.709/0.923
Oursx2	1/1	37.425/0.941	32.942/0.887	31.717/0.883	30.425/0.899

use pretrain

Method	Bit	Set5	Set14	B100	Urban100
	W/A	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
edsrx2	32/32	37.931/0.958	33.459/0.915	32.102/0.896	31.709/0.923
Oursx2	1/1	37.425/0.941	32.971/0.887	31.735/0.883	30.507/ 0.900