详解神经网络的过程

之前我们知道用卷积神经网络做图片识别

https://ieeexplore.ieee.org/document/726791/citations#citations

可是卷积神经网络是怎么工作的呢？
本文专门完整走一边神经网络，看看数据都发生了什么变化。

需要用到的库有tensorflow、matplotlib和cv2

import tensorflow as tf
import matplotlib.pyplot as plt
import cv2

读取图片

首先我们读取一张图片，然后改成320×320的尺寸

img = cv2.imread('doll.png')
b,g,r = cv2.split(img) # cv2是bgr而不是rgb，所以需要重新排列
img = cv2.merge([r,g,b])
img = cv2.resize(img, (320, 320))
img = img / 255.0

可以查看一下图片是一个320×320×3的数组

array([[[0.6627451 , 0.67843137, 0.6745098 ],
        [0.6627451 , 0.67843137, 0.6745098 ],
        [0.6627451 , 0.67843137, 0.6745098 ],
        ...,
        [0.49019608, 0.49019608, 0.48235294],
        [0.49803922, 0.48235294, 0.47843137],
        [0.49803922, 0.48235294, 0.47843137]],

       [[0.6627451 , 0.68627451, 0.67843137],
        [0.6627451 , 0.68627451, 0.67843137],
        [0.6627451 , 0.68627451, 0.67843137],
        ...,
        [0.49019608, 0.49019608, 0.48235294],
        [0.49019608, 0.48627451, 0.47843137],
        [0.49019608, 0.48627451, 0.47843137]],

       [[0.6627451 , 0.69019608, 0.69019608],
        [0.6627451 , 0.68627451, 0.68627451],
        [0.6627451 , 0.68627451, 0.68627451],
        ...,
        [0.49019608, 0.49019608, 0.48235294],
        [0.49019608, 0.48627451, 0.47843137],
        [0.49019608, 0.48627451, 0.47843137]],

       ...,
...
        [0.6627451 , 0.67843137, 0.6745098 ],
        ...,
        [0.32156863, 0.25098039, 0.20392157],
        [0.31764706, 0.24705882, 0.2       ],
        [0.31372549, 0.24313725, 0.19607843]]])

cv2的功能比较多，比如改变尺寸，如果不需要，只用自带的函数就可以完成

在matplatlib里面显示一下为

plt.imshow(img)
plt.show()