pytorch:opencv的resize和torch.nn.functional.interpolate 比较

opencv的resize和torch.nn.functional.interpolate 比较

缩放图像效果比较，
分别用resize函数和interpolate函数，输出 nearset, linear, cubic, area mode的图像
代码如下：

import cv2
import numpy as np
import torch
import torch.nn.functional as F
from matplotlib import pyplot as plt
from mmedit.core import tensor2img


def cv_scale_image(image, scale=2, interp=cv2.INTER_LINEAR):
    """
    Args:
        image: h,w,3 array
        scale
        interp:
        INTER_NEAREST - 最邻近插值
        INTER_LINEAR - 双线性插值，如果最后一个参数你不指定，默认使用这种方法
        INTER_CUBIC - 4x4像素邻域内的双立方插值
        INTER_LANCZOS4 - 8x8像素邻域内的Lanczos插值
        INTER_AREA   -
    Returns:

    """
    out = cv2.resize(image, dsize=None, fx=scale, fy=scale, interpolation=interp)
    print(out.shape, out.dtype)
    return out

def torch_scale_image(image, scale=2, interp='bilinear', align=True):
    """

    Args:
        image: h,w,3 array
        scale:
        interp: 'nearest'| 'linear'| 'bilinear'| 'bicubic'| 'trilinear'| 'area'。默认：'nearest'

    Returns:

    """
    t_image = torch.from_numpy(image / 255).permute(2, 0, 1).float().unsqueeze(0)
    # print(t_image.min(), t_image.max(), t_image.shape)
    if interp in ['linear' , 'bilinear' , 'bicubic', 'trilinear']:
        image_out = F.interpolate(
            input=t_image,
            scale_factor=scale,
            mode=interp,
            align_corners=align)
    else:
        image_out = F.interpolate(
            input=t_image,
            scale_factor=scale,
            mode=interp)

    # print(image_out.min(), image_out.max(), image_out.shape, image_out.dtype)
    return image_out
if __name__ == "__main__":
    import os
    os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"


    filename = r'D:\superresolution\realbasicvsr\RealBasicVSR-master\data\demo_000\00000000.png'

    image = cv2.imread(filename)
    image = image[..., ::-1]
    image2 = image.copy()
    # opencv resize函数
    out10 = cv_scale_image(image, scale=4, interp=cv2.INTER_NEAREST)
    out11 = cv_scale_image(image, scale=4, interp=cv2.INTER_LINEAR)
    out12 = cv_scale_image(image, scale=4, interp=cv2.INTER_CUBIC)
    out13 = cv_scale_image(image, scale=4, interp=cv2.INTER_AREA)
    print(cv2.INTER_NEAREST, cv2.INTER_LINEAR,cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_LANCZOS4)
    cv2.imwrite(filename[:-4] + 'cv' + '_nearest' + '.png', out10[..., ::-1])
    cv2.imwrite(filename[:-4] + 'cv' + '_linear' + '.png', out11[..., ::-1])
    cv2.imwrite(filename[:-4] + 'cv' + '_bicubic' + '.png', out12[..., ::-1])
    cv2.imwrite(filename[:-4] + 'cv' + '_area' + '.png', out13[..., ::-1])

    # torch interp函数
    # 'linear'| 'bilinear'| 'bicubic'| 'trilinear'| 'area'
    out20 = torch_scale_image(image2, scale=4, interp='nearest')
    out21 = torch_scale_image(image2, scale=4, interp='bilinear')
    out22 = torch_scale_image(image2, scale=4, interp='bicubic')
    out23 = torch_scale_image(image2, scale=4, interp='area')

    out20 = tensor2img(out20)
    out21 = tensor2img(out21)
    out22 = tensor2img(out22)
    out23 = tensor2img(out23)
    cv2.imwrite(filename[:-4] + 'torch_' + 'nearest.png', out20)
    cv2.imwrite(filename[:-4] + 'torch_' + 'bilinear.png', out21)
    cv2.imwrite(filename[:-4] + 'torch_' + 'bicubic.png', out22)
    cv2.imwrite(filename[:-4] + 'torch_' + 'area.png', out23)

    out21 = torch_scale_image(image2, scale=4, interp='bilinear', align=False)
    out22 = torch_scale_image(image2, scale=4, interp='bicubic', align=False)
    out21 = tensor2img(out21)
    out22 = tensor2img(out22)
    cv2.imwrite(filename[:-4] + 'torch_' + 'bilinear_align_false.png', out21)
    cv2.imwrite(filename[:-4] + 'torch_' + 'bicubic_align_false.png', out22)
    """
    align为False时, opencv和torch得到结果完全一致
    """

结论：align为False时, opencv和torch得到结果一致.
cubic图像对比结果如下：