# 1.blobFromImage
Mat cv::dnn::blobFromImage (
InputArray image,
double scalefactor = 1.0,
const Size & size = Size(),
const Scalar & mean = Scalar(),
bool swapRB = false,
bool crop = false,
int ddepth = CV_32F
)
OpenCV中的DNN模块包含blobFromImage方法对输入神经网络的图像进行处理，blobFromImage函数执行的操作及顺序为：
- 1.先相对于原图像中心resize,crop
- 2.再减均值
- 3.像素值缩放0-255 -> 0-1
- 4.图像数据通道转换，RGB->BGR
- 返回一个NCHW 数组
# 2.与torchvision对比
torchvision中可以使用transforms实现输入圣经网络图像数据的标准化，主要通过ToTensor和Normalize两个方法实现
ToTensor
- 输入为
unit8型的numpy数组或PIL.Image - 1.输入数据的
transpose,HWC->CHW - 2.将像素值范围从
0-255->0-1
Normalize
- 输入的标准化
# 3.示例
通过一段简单的程序介绍cv2.dnn.blobFromImage执行的操作与torchvision中的ToTensor+Normalize效果等同
import cv2
import torch
import torchvision
from torchvision import transforms
import numpy as np
from PIL import Image
dd = np.random.randint(0, 255, (4, 4, 3), dtype=np.uint8)
print(f"dd: {dd}")
print("===>>> by torch")
tt = torch.tensor(dd)
print(f"tt: {tt}")
tp = Image.fromarray(dd)
trans = transforms.Compose([
transforms.ToTensor(), # HWC->CHW, 0-255-> 0-1
transforms.Normalize((.5, .5, .5), (1., 1., 1.)) # (x-mwan)/std
])
trans_tt = trans(tp)
print(trans_tt)
print(f"trans_tt.shape: {trans_tt.shape}")
print("===>>> by opencv")
blob = cv2.dnn.blobFromImage(dd, 1/255, (4,4), (127.5, 127.5, 127.5), False, False)
print(f"blob: {blob}")
print(f"blog.shape: {blob.shape}")
上述代码的输出为：
tensor([[[-0.2294, 0.0412, -0.0020, -0.3667],
[ 0.2020, -0.3078, -0.4373, 0.2373],
[ 0.2333, 0.3118, 0.4255, 0.0176],
[-0.3471, 0.1235, -0.3745, 0.3196]],
[[ 0.4216, 0.3314, 0.4647, 0.4529],
[ 0.4765, 0.0451, -0.3627, 0.4098],
[ 0.3196, -0.3275, -0.4020, 0.4569],
[-0.4255, 0.4216, 0.3980, -0.2882]],
[[ 0.1314, 0.3392, -0.2098, 0.2686],
[ 0.4255, 0.3980, -0.4608, 0.0216],
[ 0.4176, 0.0294, 0.4216, -0.4961],
[-0.1745, -0.4451, 0.1941, 0.0098]]])
trans_tt.shape: torch.Size([3, 4, 4])
===>>> by opencv
blob: [[[[-0.22941178 0.04117647 -0.00196078 -0.36666667]
[ 0.2019608 -0.30784315 -0.43725494 0.23725492]
[ 0.23333335 0.31176472 0.42549023 0.01764706]
[-0.34705883 0.12352942 -0.3745098 0.31960785]]
[[ 0.42156866 0.33137256 0.4647059 0.4529412 ]
[ 0.47647062 0.04509804 -0.3627451 0.40980396]
[ 0.31960785 -0.327451 -0.40196082 0.45686278]
[-0.42549023 0.42156866 0.39803925 -0.2882353 ]]
[[ 0.13137256 0.3392157 -0.20980394 0.26862746]
[ 0.42549023 0.39803925 -0.46078435 0.02156863]
[ 0.4176471 0.02941177 0.42156866 -0.49607846]
[-0.17450981 -0.44509807 0.19411767 0.00980392]]]]
blog.shape: (1, 3, 4, 4)
可见两者的结果是一致的。