CIFAR是一个用于普通物体识别的数据集。CIFAR数据集分为两种：CIFAR-10和CIFAR-100。The CIFAR-10 and CIFAR-100 are labeled subsets of the 80 million tiny images dataset. They were collected by Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton.
CIFAR-10由60000张大小为32*32的三通道彩色图像组成，被分为10类，分别为airplane、automobile、bird、cat、deer、dog、frog、horse、ship、truck。每类由6000张图像。其中50000张图像用来训练，10000张图像用来测试。数据集分为5个训练块和1个测试块,每个块包含10000张图像.训练集每类包含5000张图像,测试集每类包含1000张图像.
CIFAR-100由60000张大小为32*32的三通道彩色图像组成，分为20个大类，每个大类又包含5个小类，总共100个小类。每个小类包含600张图像，其中500张用于训练，100张用于测试。
从https://www.cs.toronto.edu/~kriz/cifar.html 下载CIFAR C版本的二进制数据：
(1)、CIFAR-10：下载cifar-10-binary.tar.gz，解压缩，共8个文件，batches.meta.txt中存放10个种类名，data_batch_1.bin… data_batch_5.bin、test_batch.bin共6个文件，每个文件中存放10000张图像数据。
(2)、CIFAR-100：下载cifar-100-binary.tar.gz，解压缩，共5个文件，coarse_label_names.txt中存放20个大类名，fine_label_names.txt中存放100个小类名，train.bin中存放50000张训练图像，test.bin中存放10000张测试图像。
CIFAR数据集到图像转换实现的代码如下：
static void write_image_cifar(const cv::Mat& bgr, const std::string& image_save_path, const std::vector<int>& label_count, int label_class)
{
std::string str = std::to_string(label_count[label_class]);
if (label_count[label_class] < 10) {
str = "0000" + str;
} else if (label_count[label_class] < 100) {
str = "000" + str;
} else if (label_count[label_class] < 1000) {
str = "00" + str;
} else if (label_count[label_class] < 10000) {
str = "0" + str;
} else {
fprintf(stderr, "save image name fail\n");
return;
}
str = std::to_string(label_class) + "_" + str + ".png";
str = image_save_path + str;
cv::imwrite(str, bgr);
}
static void read_cifar_10(const std::string& bin_name, const std::string& image_save_path, int image_count, std::vector<int>& label_count)
{
int image_width = 32;
int image_height = 32;
std::ifstream file(bin_name, std::ios::binary);
if (file.is_open()) {
for (int i = 0; i < image_count; ++i) {
cv::Mat red = cv::Mat::zeros(image_height, image_width, CV_8UC1);
cv::Mat green = cv::Mat::zeros(image_height, image_width, CV_8UC1);
cv::Mat blue = cv::Mat::zeros(image_height, image_width, CV_8UC1);
int label_class = 0;
file.read((char*)&label_class, 1);
label_count[label_class]++;
file.read((char*)red.data, 1024);
file.read((char*)green.data, 1024);
file.read((char*)blue.data, 1024);
std::vector<cv::Mat> tmp{ blue, green, red };
cv::Mat bgr;
cv::merge(tmp, bgr);
write_image_cifar(bgr, image_save_path, label_count, label_class);
}
file.close();
}
}
int CIFAR10toImage()
{
std::string images_path = "E:/GitCode/NN_Test/data/database/CIFAR/CIFAR-10/";
// train image
std::vector<int> label_count(10, 0);
for (int i = 1; i <= 5; i++) {
std::string bin_name = images_path + "data_batch_" + std::to_string(i) + ".bin";
std::string image_save_path = "E:/GitCode/NN_Test/data/tmp/cifar-10_train/";
int image_count = 10000;
read_cifar_10(bin_name, image_save_path, image_count, label_count);
}
// test image
std::fill(&label_count[0], &label_count[0] + 10, 0);
std::string bin_name = images_path + "test_batch.bin";
std::string image_save_path = "E:/GitCode/NN_Test/data/tmp/cifar-10_test/";
int image_count = 10000;
read_cifar_10(bin_name, image_save_path, image_count, label_count);
// save big imags
images_path = "E:/GitCode/NN_Test/data/tmp/cifar-10_train/";
int width = 32 * 20;
int height = 32 * 10;
cv::Mat dst(height, width, CV_8UC3);
for (int i = 0; i < 10; i++) {
for (int j = 1; j <= 20; j++) {
int x = (j - 1) * 32;
int y = i * 32;
cv::Mat part = dst(cv::Rect(x, y, 32, 32));
std::string str = std::to_string(j);
if (j < 10)
str = "0000" + str;
else
str = "000" + str;
str = std::to_string(i) + "_" + str + ".png";
std::string input_image = images_path + str;
cv::Mat src = cv::imread(input_image, 1);
if (src.empty()) {
fprintf(stderr, "read image error: %s\n", input_image.c_str());
return -1;
}
src.copyTo(part);
}
}
std::string output_image = images_path + "result.png";
cv::imwrite(output_image, dst);
return 0;
}
static void write_image_cifar(const cv::Mat& bgr, const std::string& image_save_path,
const std::vector<std::vector<int>>& label_count, int label_class_coarse, int label_class_fine)
{
std::string str = std::to_string(label_count[label_class_coarse][label_class_fine]);
if (label_count[label_class_coarse][label_class_fine] < 10) {
str = "0000" + str;
} else if (label_count[label_class_coarse][label_class_fine] < 100) {
str = "000" + str;
} else if (label_count[label_class_coarse][label_class_fine] < 1000) {
str = "00" + str;
} else if (label_count[label_class_coarse][label_class_fine] < 10000) {
str = "0" + str;
} else {
fprintf(stderr, "save image name fail\n");
return;
}
str = std::to_string(label_class_coarse) + "_" + std::to_string(label_class_fine) + "_" + str + ".png";
str = image_save_path + str;
cv::imwrite(str, bgr);
}
static void read_cifar_100(const std::string& bin_name, const std::string& image_save_path, int image_count, std::vector<std::vector<int>>& label_count)
{
int image_width = 32;
int image_height = 32;
std::ifstream file(bin_name, std::ios::binary);
if (file.is_open()) {
for (int i = 0; i < image_count; ++i) {
cv::Mat red = cv::Mat::zeros(image_height, image_width, CV_8UC1);
cv::Mat green = cv::Mat::zeros(image_height, image_width, CV_8UC1);
cv::Mat blue = cv::Mat::zeros(image_height, image_width, CV_8UC1);
int label_class_coarse = 0;
file.read((char*)&label_class_coarse, 1);
int label_class_fine = 0;
file.read((char*)&label_class_fine, 1);
label_count[label_class_coarse][label_class_fine]++;
file.read((char*)red.data, 1024);
file.read((char*)green.data, 1024);
file.read((char*)blue.data, 1024);
std::vector<cv::Mat> tmp{ blue, green, red };
cv::Mat bgr;
cv::merge(tmp, bgr);
write_image_cifar(bgr, image_save_path, label_count, label_class_coarse, label_class_fine);
}
file.close();
}
}
int CIFAR100toImage()
{
std::string images_path = "E:/GitCode/NN_Test/data/database/CIFAR/CIFAR-100/";
// train image
std::vector<std::vector<int>> label_count;
label_count.resize(20);
for (int i = 0; i < 20; i++) {
label_count[i].resize(100);
std::fill(&label_count[i][0], &label_count[i][0] + 100, 0);
}
std::string bin_name = images_path + "train.bin";
std::string image_save_path = "E:/GitCode/NN_Test/data/tmp/cifar-100_train/";
int image_count = 50000;
read_cifar_100(bin_name, image_save_path, image_count, label_count);
// test image
for (int i = 0; i < 20; i++) {
label_count[i].resize(100);
std::fill(&label_count[i][0], &label_count[i][0] + 100, 0);
}
bin_name = images_path + "test.bin";
image_save_path = "E:/GitCode/NN_Test/data/tmp/cifar-100_test/";
image_count = 10000;
read_cifar_100(bin_name, image_save_path, image_count, label_count);
// save big imags
images_path = "E:/GitCode/NN_Test/data/tmp/cifar-100_train/";
int width = 32 * 20;
int height = 32 * 100;
cv::Mat dst(height, width, CV_8UC3);
std::vector<std::string> image_names;
for (int j = 0; j < 20; j++) {
for (int i = 0; i < 100; i++) {
std::string str1 = std::to_string(j);
std::string str2 = std::to_string(i);
std::string str = images_path + str1 + "_" + str2 + "_00001.png";
cv::Mat src = cv::imread(str, 1);
if (src.data) {
for (int t = 1; t < 21; t++) {
if (t < 10)
str = "0000" + std::to_string(t);
else
str = "000" + std::to_string(t);
str = images_path + str1 + "_" + str2 + "_" + str + ".png";
image_names.push_back(str);
}
}
}
}
for (int i = 0; i < 100; i++) {
for (int j = 0; j < 20; j++) {
int x = j * 32;
int y = i * 32;
cv::Mat part = dst(cv::Rect(x, y, 32, 32));
cv::Mat src = cv::imread(image_names[i * 20 + j], 1);
if (src.empty()) {
fprintf(stderr, "read image fail: %s\n", image_names[i * 20 + j].c_str());
return -1;
}
src.copyTo(part);
}
}
std::string output_image = images_path + "result.png";
cv::imwrite(output_image, dst);
cv::Mat src = cv::imread(output_image, 1);
if (src.empty()) {
fprintf(stderr, "read result image fail: %s\n", output_image.c_str());
return -1;
}
for (int i = 0; i < 4; i++) {
cv::Mat dst = src(cv::Rect(0, i * 800, 640, 800));
std::string str = images_path + "result_" + std::to_string(i + 1) + ".png";
cv::imwrite(str, dst);
}
return 0;
}cifar-100转换的结果如下：
