tensorflow(一)：图片处理

一、图片处理

　　1、图片存取 tf.gfile

import tensorflow as tf
import matplotlib.pyplot as plt

image_bytes = tf.gfile.FastGFile("dog.jpg", 'rb').read()  #  字节
with tf.Session() as session:
    # 2.图片解码
    img = tf.image.decode_jpeg(image_bytes)
    # print(img) # tensor('DecodePnng:0', shape=(?,?,?),dtype=uint8)
    img_array = img.eval()  # 将tensor对象转成数组
    # 3.图片显示
    plt.imshow(img_array)
    plt.show()
    # 4.图片数据类型转化(整形)
    # img = tf.image.convert_image_dtype(img, dtype=tf.float32)
    # print(img)
    # 5.图像重编码
    encode_image = tf.image.encode_jpeg(img)
    new_img = encode_image.eval()  # 数组
    # 6.图片保存
    with tf.gfile.GFile("dog_new.png", "wb") as f:
        f.write(new_img)

　　2、图片修改 tf.image

import tensorflow as tf
import matplotlib.pyplot as plt

image_bytes = tf.gfile.FastGFile("dog.jpg", 'rb').read()  #  字节
with tf.Session() as session:
    img = tf.image.decode_jpeg(image_bytes)
    # 翻转图片
    img_flipped = tf.image.flip_up_down(img)    # 上下反转
    img_flipped = tf.image.flip_left_right(img_flipped) # 左右反转
    img_flipped = tf.image.transpose_image(img_flipped) # 对角线反转
    img_flipped = tf.image.random_flip_up_down(img_flipped)    # 随机上下反转
    img_flipped = tf.image.random_flip_left_right(img_flipped) # 随机左右反转
    # 亮度设置
    img_adjust = tf.image.adjust_brightness(img_flipped, -0.5) # 增加亮度
    img_adjust = tf.image.adjust_brightness(img_adjust, +0.5)  # 降低亮度
    img_adjust = tf.image.random_brightness(img_adjust, max_delta=0.3) # 随机调整亮度，亮度在[-max_delta, +max_delta]]
    # 色度
    img_saturation = tf.image.adjust_saturation(img_adjust, 1.5)  # 支持random
　　 # 饱和度
    img_hue = tf.image.adjust_hue(img_saturation, delta=0.2)
　　 # 对比度
    img_contrast = tf.image.adjust_contrast(img_hue, 0.5)
    # 图片标准化
    img_standard = tf.image.per_image_standardization(img_adjust)
    img_standard = tf.clip_by_value(img_standard, 0.0, 10)
    # 转成数组
    img_array = img_standard.eval()
    plt.imshow(img_array)
    plt.show()

　　3、图像标注框

import tensorflow as tf
import matplotlib.pyplot as plt

image_bytes = tf.gfile.FastGFile("dog.jpg", 'rb').read()  #  字节
with tf.Session() as session:
    img = tf.image.decode_jpeg(image_bytes)
    # 调整图片大小
    img_resize = tf.image.resize_image_with_crop_or_pad(img, 300, 300)
    # 按比例截取图片
    boxes = tf.constant([[[0.31, 0.22, 0.46, 0.38], [0.38, 0.53, 0.53, 0.71]]])  # 两个标注框
    # boxes = tf.constant([[[0.31, 0.22, 0.46, 0.38]]])  # 设置一个RGB，设置四个角的比例位置
    # 给原始图片添加一个图层
    batched = tf.expand_dims(tf.image.convert_image_dtype(img_resize, tf.float32), 0)
    # 把boxes标注的框画到原始图片上
    image_with_boxes = tf.image.draw_bounding_boxes(batched, boxes)
    # 重新将原始图片设置为RGB
    image_with_boxes = tf.reshape(image_with_boxes, [300, 300, 3])
    img_array = image_with_boxes.eval()
    plt.imshow(img_array)
    plt.show()

　　4、图片随机截取

import matplotlib.pyplot as plt

image_bytes = tf.gfile.FastGFile("dog.jpg", 'rb').read()  #  字节
with tf.Session() as session:
    img = tf.image.decode_jpeg(image_bytes)
    # 给定截取框大小
    bounding_boxes = tf.constant([[[0.31, 0.22, 0.46, 0.38]]])  # 设置一个RGB，设置四个角的比例位置
    # 选择相关图像截取算法截图
    # Bounding boxes are supplied and returned as `[y_min, x_min, y_max, x_max]`.
    begin, size, bboxes = tf.image.sample_distorted_bounding_box(
        tf.shape(img), bounding_boxes=bounding_boxes, min_object_covered=0.1
    )
    # 生成概要
    # img_with_box = tf.image.draw_bounding_boxes(tf.expand_dims(tf.image.convert_image_dtype(img, dtype=tf.float32), 0), bboxes)
    # tf.summary.image('img_with_box', img_with_box)
    # print(begin.eval(), size.eval())
    # 截图
    distorted_img = tf.slice(img, begin, size)
    img_array = distorted_img.eval()
    plt.imshow(img_array)
    plt.show()

　　5、一个简单样例代码，实现随机截取图片

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

class Sample:
    def load_jpg(self, path, mode='rb'):
        image_bytes = tf.gfile.FastGFile(path, mode).read()
        return tf.image.decode_jpeg(image_bytes, channels=3)
    def _distort_picture(self, image, color_ordering=0):
        if color_ordering == 0:
            image = tf.image.random_brightness(image, max_delta=32./255.)  # 随机亮度
            image = tf.image.random_contrast(image, lower=0.5, upper=1.5)  # 对比度
            image = tf.image.random_hue(image, max_delta=0.2)              # 饱和度
            image = tf.image.random_saturation(image, lower=0.5, upper=1.5)# 色度
        if color_ordering == 1:
            image = tf.image.random_hue(image, max_delta=0.2)              # 饱和度
            image = tf.image.random_saturation(image, lower=0.5, upper=1.5)# 色度
            image = tf.image.random_flip_left_right(image)
            image = tf.image.random_flip_up_down(image)
        return tf.clip_by_value(image, clip_value_min=0.0, clip_value_max=1.0)  # 归一化
    def _preprocess_for_train(self, image, height, width, bounding_boxes=None):
        if bounding_boxes is None:
            bounding_boxes = tf.constant([0.0, 0.0, 1.0, 1.0], dtype=tf.float32, shape=[1, 1, 4])
        if image.dtype != tf.float32:
            image = tf.image.convert_image_dtype(image, dtype=tf.float32)
        begin, size, bboxes = tf.image.sample_distorted_bounding_box(
            tf.shape(image),  bounding_boxes=bounding_boxes, min_object_covered=0.1
        )
        # 随机截图
        distorted_image = tf.slice(image, begin=begin, size=size)
        # 调整随机截图的图片大小
        # distorted_image = tf.image.resize_image_with_crop_or_pad(distorted_image, height, width)
        distorted_image = tf.image.resize_images(
            distorted_image, size=[height, width], method=np.random.randint(4)
        )
        # 随机调整图片的一些设置
        distorted_image = self._distort_picture(distorted_image, np.random.randint(2))
        return distorted_image
    def get_random_picture(self, number, image, *args, **kwargs):
        with tf.Session() as session:
            for i in range(number):
                random_picture = self._preprocess_for_train(image, *args, **kwargs)
                plt.imshow(random_picture.eval())
                plt.show()

def main():
    sample = Sample()
    image = sample.load_jpg("dog.jpg", 'rb')
#     bounding_boxes = tf.constant([0.2, 0.2, 0.8, 0.8], dtype=tf.float32, shape=[1, 1, 4])
    bounding_boxes = tf.constant([[[0.2, 0.2, 0.8, 0.8]]])
    height = width = 150
    sample.get_random_picture(5, image, height, width, bounding_boxes)
main()

　　5、图片处理有关函数整理

函数	描述
tf.gfile.FastGFile	读取单个图片，返回字节流数据
tf.decode_jpeg	在图片读入操作之后，图片处理之前，对图片进行解码
tf.encode_jpeg	在图片保存时对图片进行重编码
tf.gfile.GFile	写出单个图片
tf.image.convert_image_dtype	转换图片的数据类型
tf.resize_images	剪裁图片大小
tf.resize_image_with_crop_of_pad	剪裁单个图片大小
tf.image.random_flip_left_right	图片随机左右反转
tf.image.random_flip_up_down	图片随机上下反转
tf.image.random_brightness	图片随机调整亮度
tf.image.random_hue	图片随机调整饱和度
tf.image.random_contrast	图片随机调整对比度
tf.image.random_saturation	图片随机调整色度
tf.image.per_image_standardization	单个图片标准化
tf.image.clip_by_value	单个图片归一化，其它还有tf.image.clip_by_XXX等方法
tf.expand_dims	给图片增加维度(图层)
tf.image.sample_distorted_bounding_box	生成随机子图
tf.image.draw_bounding_boxes	将标注框标注的子图取出来
tf.image.reshape	调整图片的维度
tf.slice	截取随机子图为单个图片

二、TFRecord

　　TFRecord文件是tensorflow指定的一种文件存储格式。它由tf.train.Example和tf.train.Feature来规定和实现。

# tf.train.Example Protocol Buffer
message Example {
    Features features = 1;
}
message Features {
    map<string, Feature> feature = 1;
}
message Feature {
    oneof kind{
        BytesList bytes_list = 1;
        FloatList float_list = 2;
        Int64List int64_list = 3;
    }
}

　　1、TFRecord文件写出

　　手写字mnist数据下载地址: http://yann.lecun.com/exdb/mnist/

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np

# 导入训练集和测试集的图片和标签
mnist = input_data.read_data_sets("tensorflow/mnist/", dtype=tf.uint8, one_hot=True)
# 获取图片和标签
images = mnist.train.images  # images.shape (55000, 784) 热独编码
labels = mnist.train.labels  # labels.shape (55000, 10)
# 获取图像的数量及图片的分辨率([......])
numbers, pixels = images.shape

# 按照tf.train.Example Protocol Buffer来定义TFRecord文件格式
def _int64(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))
def _bytes(value):
    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
def example_protocol_buffer(pixel, size, image):
    example = tf.train.Example(
        features=tf.train.Features(
            feature={
                'pixels': _int64(pixels),
                'label': _int64(size),
                'image': _bytes(image)
            }
        )
    )
    return example.SerializeToString()  # 序列化为字节

# 输出文件地址
filename = "tensorflow/test/mnist.tfrecord"
# 创建一个writer
writer = tf.python_io.TFRecordWriter(filename)
# 遍历每张图片
for index in range(numbers):
    image = images[index].tostring()  # 转成字节
    serialize = example_protocol_buffer(pixels, np.argmax(labels[index]), image)
    writer.write(serialize)
writer.close()
print("done.")

　　2、TFRecord文件读入

import tensorflow as tf
import matplotlib.pyplot as plt

# 创建reader
reader = tf.TFRecordReader()
# 创建字节流读取队列
filename_queue = tf.train.string_input_producer(
    ["tensorflow/test/mnist.tfrecord"]
)
# 从文件中读取一个样例，read_up_to函数一次性读取多个样例
key, serialized_example = reader.read(filename_queue)
# 解析读取的一个样例，如果需要解析多个样例，可以用parse_example
def parse_single(serialized_example):
    features = tf.parse_single_example(
        serialized_example,
        features={
            'image': tf.FixedLenFeature([], tf.string),
            'label': tf.FixedLenFeature([], tf.int64),
            'pixels': tf.FixedLenFeature([], tf.int64)
        }
    )
    # 将读取的单个样例解码
    image = tf.decode_raw(features['image'], tf.uint8)
    label = tf.cast(features['label'], tf.int32)
    pixels = tf.cast(features['pixels'], tf.int32)
    return image, label, pixels

sess = tf.Session()
# 启动多线程处理输入数据
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)

for i in range(10):
    image, label, pixel = sess.run(parse_single(serialized_example))
    print(image, label, pixel)
    plt.imshow(image.reshape(28, 28))
    plt.show()