app.py

import numpy as np
import gradio as gr
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torchvision import transforms
from PIL import Image


title = "Super Resolution with CNN"
description = """

Your low resolution image will be reconstructed to high resolution with a scale of 2 with a convolutional neural network!<br>

Detailed training and dataset can be found on my [github repo](https://github.com/susuhu/super-resolution).<br>

"""

article = """
<div style='margin:20px auto;'>
<p>Sources:<p>
<p>📜 <a href="https://arxiv.org/abs/1501.00092">Image Super-Resolution Using Deep Convolutional Networks</a></p>
<p>📦 Dataset <a href="https://github.com/eugenesiow/super-image-data">this GitHub repo</a></p>
</div>
"""
examples = [
    ["LR_image.png"],
    ["barbara.png"],
]


class SRCNNModel(nn.Module):
    def __init__(self):
        super(SRCNNModel, self).__init__()
        self.conv1 = nn.Conv2d(1, 64, 9, padding=4)
        self.conv2 = nn.Conv2d(64, 32, 1, padding=0)
        self.conv3 = nn.Conv2d(32, 1, 5, padding=2)

    def forward(self, x):
        out = F.relu(self.conv1(x))
        out = F.relu(self.conv2(out))
        out = self.conv3(out)
        return out


def pred_SRCNN(model, image, device, scale_factor=2):
    """
    model: SRCNN model
    image: low resolution image PILLOW image
    scale_factor: scale factor for resolution
    device: cuda or cpu
    """
    model.to(device)
    model.eval()

    # open image, gradio opens image as nparray
    image = Image.fromarray(image)
    # split channels
    y, cb, cr = image.convert("YCbCr").split()
    # size will be used in image transform
    original_size = y.size

    # bicubic interpolate it to the original size
    y_bicubic = transforms.Resize(
        (original_size[1] * scale_factor, original_size[0] * scale_factor),
        interpolation=transforms.InterpolationMode.BICUBIC,
    )(y)
    cb_bicubic = transforms.Resize(
        (original_size[1] * scale_factor, original_size[0] * scale_factor),
        interpolation=transforms.InterpolationMode.BICUBIC,
    )(cb)
    cr_bicubic = transforms.Resize(
        (original_size[1] * scale_factor, original_size[0] * scale_factor),
        interpolation=transforms.InterpolationMode.BICUBIC,
    )(cr)
    # turn it into tensor and add batch dimension
    y_bicubic = transforms.ToTensor()(y_bicubic).to(device).unsqueeze(0)
    # get the y channel SRCNN prediction
    y_pred = model(y_bicubic)
    # convert it to numpy image
    y_pred = y_pred[0].cpu().detach().numpy()

    # convert it into regular image pixel values
    y_pred = y_pred * 255
    y_pred.clip(0, 255)
    # conver y channel from array to PIL image format for merging
    y_pred_PIL = Image.fromarray(np.uint8(y_pred[0]), mode="L")
    # merge the SRCNN y channel with cb cr channels
    out_final = Image.merge("YCbCr", [y_pred_PIL, cb_bicubic, cr_bicubic]).convert(
        "RGB"
    )

    image_bicubic = transforms.Resize(
        (original_size[1] * scale_factor, original_size[0] * scale_factor),
        interpolation=transforms.InterpolationMode.BICUBIC,
    )(image)
    return out_final, image_bicubic, image


# load model
# print("Loading  SRCNN model...")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = SRCNNModel().to(device)
model.load_state_dict(
    torch.load("SRCNNmodel_trained.pt", map_location=torch.device(device))
)
model.eval()
# print("SRCNN model loaded!")


# def image_grid(imgs, rows, cols):
#     '''
#     imgs:list of PILImage
#     '''
#     assert len(imgs) == rows*cols

#     w, h = imgs[0].size
#     grid = Image.new('RGB', size=(cols*w, rows*h))
#     grid_w, grid_h = grid.size

#     for i, img in enumerate(imgs):
#         grid.paste(img, box=(i%cols*w, i//cols*h))
#     return grid


def super_reso(image):
    # gradio open image as np array
    #image_array = np.asarray(image_path)
    #image = Image.fromarray(image_array, mode="RGB") 

    # prediction
    with torch.no_grad():
        out_final, image_bicubic, image = pred_SRCNN(
            model=model, image=image, device=device
        )
    # grid = image_grid([out_final,image_bicubic],1,2)
    return out_final, image_bicubic


gr.Interface(
    fn=super_reso,
    inputs=gr.inputs.Image(label="Upload image"),
    outputs=[
        gr.outputs.Image(label="Convolutional neural network"),
        gr.outputs.Image(label="Bicubic interpoloation"),
    ],
    title=title,
    description=description,
    article=article,
    examples=examples,
).launch()