han

HAN

Input

(Image from https://github.com/sanghyun-son/EDSR-PyTorch/blob/master/test/0853x4.png)

Ailia input shape : (1, 3, IMAGE_HEIGHT, IMAGE_WIDTH)

Output

Ailia output shape : (1, 3, IMAGE_HEIGHT * scale, IMAGE_WIDTH * scale)

default : scale=2

Usage

Automatically downloads the onnx and prototxt files when running. It is necessary to be connected to the Internet while downloading.

For the sample image with twice the resolution (BI),

$ python3 han.py

If you want to specify the input image, put the image path after the --input option.
You can use --savepath option to change the name of the output file to save.

$ python3 han.py --input IMAGE_PATH --savepath SAVE_IMAGE_PATH

If you want to specify the scale for the resolution, put the scale after the --scale option.
Choose the scale in [2, 3, 4, 8].

$ python3 han.py --scale SCALE 

If you want to the model trained on imaged degraded by the Blur-downscale Degradation Model (BD), specify the --blur option.
Only a 3-resolution scale can be used with this option.

$ python3 han.py --scale 3 --blur 

By adding the --video option, you can input the video.
If you pass 0 as an argument to VIDEO_PATH, you can use the webcam input instead of the video file.

$ python3 han.py --video VIDEO_PATH

The default setting is to use the optimized model and weights, but you can also switch to the normal model by using the --normal option. If the output image is entirely black, try to add the -e 0 option.

$ python3 han.py -e 0

Reference

Single Image Super-Resolution via a Holistic Attention Network

Framework

Pytorch 1.3.0

Model Format

ONNX opset = 11

Netron

han_BIX2.onnx.prototxt han_BIX2.opt.onnx.prototxt

han_BIX3.onnx.prototxt han_BIX3.opt.onnx.prototxt

han_BIX4.onnx.prototxt han_BIX4.opt.onnx.prototxt

han_BIX8.onnx.prototxt han_BIX8.opt.onnx.prototxt

han_BDX3.onnx.prototxt han_BDX3.opt.onnx.prototxt