想多看看OpenVINO 2021r2对各种SR模型计算层的支持怎么样，再试一下DBPN, 这个是NTIRE2018超分8倍的冠军， 只随便测试一下2X的模型
从网上找了个基于caffe的模型 https://github.com/alterzero/DBPN-caffe caffe是个非常老的框架了 OpenVINO应该支持的非常好了.
关于推理数据的预处理和后处理，从原项目可以看出就是简单的除以255.0和乘以255.0
%% work on illuminance only
im_gnd = modcrop(im, up_scale);
输入低分辨率图像就是im_gnd除以255然后缩小up_scale倍
im_gnd = single(im_gnd)/255;
im_l = imresize(im_gnd, 1/up_scale, 'bicubic');
%% bicubic interpolation
tic
im_b = imresize(im_l, up_scale, 'bicubic');
t_bic=toc;
%% DBPN
if flat==1
[im_dbpn,t_dbpn] = run_cnn(im_b, model, weights,use_gpu);
t_dbpn=t_dbpn+t_bic;
else
if ensemble==0
[im_dbpn,t_dbpn] = run_cnn(im_l, model, weights,use_gpu);
else
[im_dbpn,t_dbpn]=selfEnsemble(im_l,model,weights,use_gpu);
end
end
%% remove border
输出数据乘以255
im_dbpn = shave(uint8(im_dbpn * 255), [up_scale, up_scale]);
im_gnd = shave(uint8(im_gnd * 255), [up_scale, up_scale]);
im_b = shave(uint8(im_b * 255), [up_scale, up_scale]);
废话不多说，MO转换大法
python "c:\Program Files (x86)\IntelSWTools\openvino_2021\deployment_tools\model_optimizer\mo_caffe.py" --input_shape=[1,3,480,640] --scale_values=[255.0] --input_model=DBPN_2x.caffemodel --data_type FP16
后处理部分针对推理输出数据做一下乘255.0的处理
//channel == 3
cv::Mat Img(H, W, CV_8UC3);
unsigned char *image_ptr = Img.data;
for (size_t n = 0; n < num_images; n++) {
for (size_t i = 0; i < nPixels; i++) {
//输出数据乘以255.0
data_img[i * num_channels] = static_cast<float>(output_data[i + n * nPixels * num_channels]*255.0);
data_img[i * num_channels + 1] = static_cast<float>(
output_data[(i + nPixels) + n * nPixels * num_channels] * 255.0 );
data_img[i * num_channels + 2] = static_cast<float>(
output_data[(i + 2 * nPixels) + n * nPixels * num_channels] * 255.0);
//switch BGR->RGB, OpenCV doesn't need it, just skip it
//float temp = data_img[i * num_channels];
//data_img[i * num_channels] = data_img[i * num_channels + 2];
//data_img[i * num_channels + 2] = temp;
if (data_img[i * num_channels] < 0) data_img[i * num_channels] = 0;
if (data_img[i * num_channels] > 255) data_img[i * num_channels] = 255;
image_ptr[i * num_channels] = data_img[i * num_channels];
if (data_img[i * num_channels + 1] < 0) data_img[i * num_channels + 1] = 0;
if (data_img[i * num_channels + 1] > 255) data_img[i * num_channels + 1] = 255;
image_ptr[i * num_channels + 1] = data_img[i * num_channels + 1];
if (data_img[i * num_channels + 2] < 0) data_img[i * num_channels + 2] = 0;
if (data_img[i * num_channels + 2] > 255) data_img[i * num_channels + 2] = 255;
image_ptr[i * num_channels + 2] = data_img[i * num_channels + 2];
}
}
imshow("DBPN_2x", Img);
最终得到输出效果
原始图片（测试图片来自网络）
Bicubic的2x放大效果
DBPN 2X效果
肉眼看上去好像比EDSR要好一点 :)
最终照例看一下性能，
调用inferRequest_regular.Infer()推理的时间, 在8665U 4核8线程的CPU和 Gen9 24EU的核显上
CPU: 23069ms (0.043FPS)
GPU: 13147ms (0.076FPS)
看来OpenVINO做超分的主要难度在MO转换上，只要MO能够正常的转换成功，后面的推理代码实现基本就是一个OpenVINO的async_classification_sample例程的代码实现。对于2021年发布的OpenVINO版本来说，转换这些前几年的SR模型，就像用机关枪打冷兵器时代的战争一样，完全属于*，轻松畅快 :)
实现代码，仅供参考
https://gitee.com/tisandman/dbpn_ov2021
————————————————
版权声明：本文为CSDN博主「sandmangu」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/sandmangu/article/details/113508669