OpenCV 4.12.0
开源计算机视觉
加载中...
搜索中...
无匹配项
超分辨率基准测试

基准测试

超分辨率模块包含用于基准测试的示例代码，以比较不同的模型和算法。此处提供了一个执行基准测试的示例代码，并收集了一些基准测试结果。测试是在运行 Ubuntu 18.04.02 操作系统的 Intel i7-9700K CPU 上进行的。

示例源代码

1// This file is part of OpenCV project.
2// It is subject to the license terms in the LICENSE file found in the top-level directory
3// of this distribution and at https://opencv.ac.cn/license.html.
4
5#include <iostream>
6#include <opencv2/opencv_modules.hpp>
7
8#ifdef HAVE_OPENCV_QUALITY
9#include <opencv2/dnn_superres.hpp>
10#include <opencv2/quality.hpp>
11#include <opencv2/imgproc.hpp>
12#include <opencv2/highgui.hpp>
13
14using namespace std;
15using namespace cv;
16using namespace dnn_superres;
17
18static void showBenchmark(vector<Mat> images, string title, Size imageSize,
19 const vector<String> imageTitles,
20 const vector<double> psnrValues,
21 const vector<double> ssimValues)
22{
23 int fontFace = FONT_HERSHEY_COMPLEX_SMALL;
24 int fontScale = 1;
25 Scalar fontColor = Scalar(255, 255, 255);
26
27 int len = static_cast<int>(images.size());
28
29 int cols = 2, rows = 2;
30
31 Mat fullImage = Mat::zeros(Size((cols * 10) + imageSize.width * cols, (rows * 10) + imageSize.height * rows),
32 images[0].type());
33
34 stringstream ss;
35 int h_ = -1;
36 for (int i = 0; i < len; i++) {
37
38 int fontStart = 15;
39 int w_ = i % cols;
40 if (i % cols == 0)
41 h_++;
42
43 Rect ROI((w_ * (10 + imageSize.width)), (h_ * (10 + imageSize.height)), imageSize.width, imageSize.height);
44 Mat tmp;
45 resize(images[i], tmp, Size(ROI.width, ROI.height));
46
47 ss << imageTitles[i];
48 putText(tmp,
49 ss.str(),
50 Point(5, fontStart),
51 fontFace,
52 fontScale,
53 fontColor,
54 1,
55 16);
56
57 ss.str("");
58 fontStart += 20;
59
60 ss << "PSNR: " << psnrValues[i];
61 putText(tmp,
62 ss.str(),
63 Point(5, fontStart),
64 fontFace,
65 fontScale,
66 fontColor,
67 1,
68 16);
69
70 ss.str("");
71 fontStart += 20;
72
73 ss << "SSIM: " << ssimValues[i];
74 putText(tmp,
75 ss.str(),
76 Point(5, fontStart),
77 fontFace,
78 fontScale,
79 fontColor,
80 1,
81 16);
82
83 ss.str("");
84 fontStart += 20;
85
86 tmp.copyTo(fullImage(ROI));
87 }
88
89 namedWindow(title, 1);
90 imshow(title, fullImage);
91 waitKey();
92}
93
94static Vec2d getQualityValues(Mat orig, Mat upsampled)
95{
96 double psnr = PSNR(upsampled, orig);
97 Scalar q = quality::QualitySSIM::compute(upsampled, orig, noArray());
98 double ssim = mean(Vec3d((q[0]), q[1], q[2]))[0];
99 return Vec2d(psnr, ssim);
100}
101
102int main(int argc, char *argv[])
103{
104 // Check for valid command line arguments, print usage
105 // if insufficient arguments were given.
106 if (argc < 4) {
107 cout << "用法：参数 1：图像路径 | 图像路径" << endl;
108 cout << "\t 参数 2：算法 | edsr, espcn, fsrcnn 或 lapsrn" << endl;
109 cout << "\t 参数 3：模型文件路径 2 \n";
110 cout << "\t 参数 4：缩放比例 | 2, 3, 4 或 8 \n";
111 return -1;
112 }
113
114 string path = string(argv[1]);
115 string algorithm = string(argv[2]);
116 string model = string(argv[3]);
117 int scale = atoi(argv[4]);
118
119 Mat img = imread(path);
120 if (img.empty()) {
121 cerr << "无法加载图像：" << img << "\n";
122 return -2;
123 }
124
125 //Crop the image so the images will be aligned
126 int width = img.cols - (img.cols % scale);
127 int height = img.rows - (img.rows % scale);
128 Mat cropped = img(Rect(0, 0, width, height));
129
130 //Downscale the image for benchmarking
131 Mat img_downscaled;
132 resize(cropped, img_downscaled, Size(), 1.0 / scale, 1.0 / scale);
133
134 //Make dnn super resolution instance
135 DnnSuperResImpl sr;
136
137 vector <Mat> allImages;
138 Mat img_new;
139
140 //Read and set the dnn model
141 sr.readModel(model);
142 sr.setModel(algorithm, scale);
143 sr.upsample(img_downscaled, img_new);
144
145 vector<double> psnrValues = vector<double>();
146 vector<double> ssimValues = vector<double>();
147
148 //DL MODEL
149 Vec2f quality = getQualityValues(cropped, img_new);
150
151 psnrValues.push_back(quality[0]);
152 ssimValues.push_back(quality[1]);
153
154 cout << sr.getAlgorithm() << "：" << endl;
155 cout << "PSNR：" << quality[0] << " SSIM：" << quality[1] << endl;
156 cout << "----------------------" << endl;
157
158 //BICUBIC
159 Mat bicubic;
160 resize(img_downscaled, bicubic, Size(), scale, scale, INTER_CUBIC);
161 quality = getQualityValues(cropped, bicubic);
162
163 psnrValues.push_back(quality[0]);
164 ssimValues.push_back(quality[1]);
165
166 cout << "双三次插值" << endl;
167 cout << "PSNR：" << quality[0] << " SSIM：" << quality[1] << endl;
168 cout << "----------------------" << endl;
169
170 //NEAREST NEIGHBOR
171 Mat nearest;
172 resize(img_downscaled, nearest, Size(), scale, scale, INTER_NEAREST);
173 quality = getQualityValues(cropped, nearest);
174
175 psnrValues.push_back(quality[0]);
176 ssimValues.push_back(quality[1]);
177
178 cout << "最近邻插值" << endl;
179 cout << "PSNR：" << quality[0] << " SSIM：" << quality[1] << endl;
180 cout << "----------------------" << endl;
181
182 //LANCZOS
183 Mat lanczos;
184 resize(img_downscaled, lanczos, Size(), scale, scale, INTER_LANCZOS4);
185 quality = getQualityValues(cropped, lanczos);
186
187 psnrValues.push_back(quality[0]);
188 ssimValues.push_back(quality[1]);
189
190 cout << "Lanczos" << endl;
191 cout << "PSNR：" << quality[0] << " SSIM：" << quality[1] << endl;
192 cout << "-----------------------------------------------" << endl;
193
194 vector <Mat> imgs{img_new, bicubic, nearest, lanczos};
195 vector <String> titles{sr.getAlgorithm(), "Bicubic", "Nearest neighbor", "Lanczos"};
196 showBenchmark(imgs, "质量基准测试", Size(bicubic.cols, bicubic.rows), titles, psnrValues, ssimValues);
197
198 waitKey(0);
199
200 return 0;
201}
202#else
203int main()
204{
205 std::cout << "此示例需要 OpenCV Quality 模块。" << std::endl;
206 return 0;
207}
208#endif
cv::Mat
n 维密集数组类
定义 mat.hpp:830
cv::Mat::copyTo
void copyTo(OutputArray m) const
将矩阵复制到另一个矩阵。
cv::Mat::cols
int cols
定义 mat.hpp:2165
定义 interface.h:61
cv::getTickFrequency
double getTickFrequency()
cv::Mat::rows
int rows
行数和列数，或者当矩阵维度超过 2 时为 (-1, -1)
定义 mat.hpp:2165
cv::Point_< int >
cv::Rect_
2D 矩形的模板类。
定义 types.hpp:444
cv::Scalar_< double >
cv::Size_
用于指定图像或矩形大小的模板类。
Definition types.hpp:335
cv::Size_::height
_Tp height
高度
Definition types.hpp:363
cv::Size_::width
_Tp width
宽度
Definition types.hpp:362
dnn_superres.hpp
cv::waitKey
int waitKey(int delay=0)
等待按键按下。
cv::imread
CV_EXPORTS_W Mat imread(const String &filename, int flags=IMREAD_COLOR_BGR)
从文件加载图像。
cv::resize
void resize(InputArray src, OutputArray dst, Size dsize, double fx=0, double fy=0, int interpolation=INTER_LINEAR)
调整图像大小。
main
int main(int argc, char *argv[])
定义 highgui_qt.cpp:3
cv::quality::quality_utils::scale
void scale(cv::Mat &mat, const cv::Mat &range, const T min, const T max)
Definition quality_utils.hpp:90
cv
定义 core.hpp:107
std
STL 命名空间。

解释

  1. 读取并缩小图像

    int width = img.cols - (img.cols % scale);
    int height = img.rows - (img.rows % scale);
    Mat cropped = img(Rect(0, 0, width, height));
    Mat img_downscaled;
    cv::resize(cropped, img_downscaled, cv::Size(), 1.0 / scale, 1.0 / scale);

    按缩放因子调整图像大小。在此之前需要进行裁剪，以便图像对齐。

  2. 设置模型

    DnnSuperResImpl sr;
    sr.readModel(path);
    sr.setModel(algorithm, scale);
    sr.upsample(img_downscaled, img_new);

    实例化一个 DNN 超分辨率对象。读取并设置算法和缩放因子。

  3. 执行基准测试

    double psnr = PSNR(img_new, cropped);
    Scalar q = cv::quality::QualitySSIM::compute(img_new, cropped, cv::noArray());
    double ssim = mean(cv::Vec3f(q[0], q[1], q[2]))[0];
    cv::quality::QualitySSIM::compute
    cv::Scalar compute(InputArray cmp) CV_OVERRIDE
    计算 SSIM。
    返回一个空的 InputArray 或 OutputArray。
    InputOutputArray noArray()
    template<typename _Tp , int m, int n>

    计算 PSNR 和 SSIM。使用 OpenCV 的 PSNR（核心 OpenCV）和 SSIM（contrib 模块）函数来比较图像。对其他上采样算法（例如其他深度学习模型或插值方法，如双三次插值、最近邻插值）重复此过程。

基准测试结果

General100 数据集

2 倍缩放因子

平均推理时间（秒，CPU）平均 PSNR平均 SSIM
ESPCN0.008795 32.7059 0.9276
EDSR5.923450 34.1300 0.9447
FSRCNN0.021741 32.8886 0.9301
LapSRN0.114812 32.2681 0.9248
Bicubic0.000208 32.1638 0.9305
Nearest neighbor0.000114 29.1665 0.9049
Lanczos0.001094 32.4687 0.9327

3 倍缩放因子

平均推理时间（秒，CPU）平均 PSNR平均 SSIM
ESPCN0.005495 28.4229 0.8474
EDSR2.455510 29.9828 0.8801
FSRCNN0.008807 28.3068 0.8429
LapSRN0.282575 26.7330 0.8862
Bicubic0.000311 26.0635 0.8754
Nearest neighbor0.000148 23.5628 0.8174
Lanczos0.001012 25.9115 0.8706

4 倍缩放因子

平均推理时间（秒，CPU）平均 PSNR平均 SSIM
ESPCN0.004311 26.6870 0.7891
EDSR1.607570 28.1552 0.8317
FSRCNN0.005302 26.6088 0.7863
LapSRN0.121229 26.7383 0.7896
Bicubic0.000311 26.0635 0.8754
Nearest neighbor0.000148 23.5628 0.8174
Lanczos0.001012 25.9115 0.8706

图像

2 倍缩放因子

Set5: butterfly.png尺寸: 256x256

PSNR / SSIM / 速度 (CPU)26.6645 / 0.9048 / 0.000201 23.6854 / 0.8698 / 0.000075 26.9476 / 0.9075 / 0.001039

29.0341 / 0.9354 / 0.004157 29.0077 / 0.9345 / 0.006325 27.8212 / 0.9230 / 0.037937 30.0347 / 0.9453 / 2.077280

3 倍缩放因子

Urban100: img_001.png尺寸: 1024x644

PSNR / SSIM / 速度 (CPU)27.0474 / 0.8484 / 0.000391 26.0842 / 0.8353 / 0.000236 27.0704 / 0.8483 / 0.002234

LapSRN 未针对 3 倍进行训练
由于其架构原因

28.0118 / 0.8588 / 0.030748 28.0184 / 0.8597 / 0.094173 30.5671 / 0.9019 / 9.517580

4 倍缩放因子

Set14: comic.png尺寸: 250x361

PSNR / SSIM / 速度 (CPU)19.6766 / 0.6413 / 0.000262 18.5106 / 0.5879 / 0.000085 19.4948 / 0.6317 / 0.001098

20.0417 / 0.6302 / 0.001894 20.0885 / 0.6384 / 0.002103 20.0676 / 0.6339 / 0.061640 20.5233 / 0.6901 / 0.665876

8 倍缩放因子

Div2K: 0006.png尺寸: 1356x2040

PSNR / SSIM / 速度 (CPU)26.3139 / 0.8033 / 0.001107 23.8291 / 0.7340 / 0.000611

26.1565 / 0.7962 / 0.004782 26.7046 / 0.7987 / 2.274290
由 Doxygen 于 2025 年 7 月 3 日星期四 12:14:36 为 OpenCV 生成 doxygen 1.12.0