opencv4.5.x+cuda11.x和源码编译及yolov7用C++DNN部署测试

一、准备工作

1、下载opencv源码,下载OpenCV4.5.4
https://github.com/opencv/opencv

https://github.com/opencv/opencv_contrib
同理opencv_contrib也下载4.5.4版本，点击tags,找到4.5.4版本，点击进去，下载如下的压缩包解压。

2、下载cmake，安装版本为：cmake 3.13.4

https://cmake.org/

二、CMake编译工作

3、新建cudabuild

点击config/选择编译版本对应64位的
4.

ffmpeg:
下载opencv_videoio_ffmpeg_64.dll和opencv_videoio_ffmpeg.dll放E:\opencv\opencv\sources\3rdparty\ffmpeg目录下
https://raw.githubusercontent.com/opencv/opencv_3rdparty/40b4666d1aa374205fd61373496e15d92ecd5313/ffmpeg/opencv_videoio_ffmpeg_64.dll
https://raw.githubusercontent.com/opencv/opencv_3rdparty/40b4666d1aa374205fd61373496e15d92ecd5313/ffmpeg/opencv_videoio_ffmpeg.dll
添加哈希code的dll文件放E:\opencv\opencv\sources\.cache\ffmpeg

ippicv同理：
打开cmake文件手动下载


添加完下载后，点击generate生成完毕。
之后，添加扩展包路径：
再者cuda支持都勾上；

然后，test包取消编译；

以上都完成后再次点击config,生成完成后有个CUDA_FAST_MATH勾上，它上一个不要勾

然后取消xfeature2d模块，不要打勾

同时取消wechat_qrcode模块，不要打勾

最后取消face模块，不要打勾

最后点击generate，编译完成。

三、vs配置前生成libopencv库

一、E:\opencv\opencv\cudabuild下CMake编译完成后，会生成OpenCV.sin点击打开进入vs，选择release/x64
点击ALL_BUILD右键生成。

点击INSTALL右键生成。

四、vs配置生成的libopencv库（一次配置后续不用配）

关闭解决方案，点击视图/其他窗口/属性管理器/releasex64/Microsoft.Cpp.x64.user/属性
添加包含目录:

E:\opencv\opencv\cudabuild\install\include\opencv2
E:\opencv\opencv\cudabuild\install\include

添加库目录:

E:\opencv\opencv\cudabuild\install\x64\vc15\lib

连接器/输入/附加依赖项

opencv_aruco454.lib
opencv_barcode454.lib
opencv_bgsegm454.lib
opencv_calib3d454.lib
opencv_ccalib454.lib
opencv_core454.lib
opencv_cudaarithm454.lib
opencv_cudabgsegm454.lib
opencv_cudacodec454.lib
opencv_cudafeatures2d454.lib
opencv_cudafilters454.lib
opencv_cudaimgproc454.lib
opencv_cudalegacy454.lib
opencv_cudaobjdetect454.lib
opencv_cudaoptflow454.lib
opencv_cudastereo454.lib
opencv_cudawarping454.lib
opencv_cudev454.lib
opencv_datasets454.lib
opencv_dnn454.lib
opencv_dnn_objdetect454.lib
opencv_dnn_superres454.lib
opencv_dpm454.lib
opencv_features2d454.lib
opencv_flann454.lib
opencv_fuzzy454.lib
opencv_hdf454.lib
opencv_hfs454.lib
opencv_highgui454.lib
opencv_imgcodecs454.lib
opencv_imgproc454.lib
opencv_img_hash454.lib
opencv_intensity_transform454.lib
opencv_line_descriptor454.lib
opencv_mcc454.lib
opencv_ml454.lib
opencv_objdetect454.lib
opencv_optflow454.lib
opencv_phase_unwrapping454.lib
opencv_photo454.lib
opencv_plot454.lib
opencv_quality454.lib
opencv_rapid454.lib
opencv_reg454.lib
opencv_rgbd454.lib
opencv_saliency454.lib
opencv_shape454.lib
opencv_stereo454.lib
opencv_stitching454.lib
opencv_structured_light454.lib
opencv_superres454.lib
opencv_surface_matching454.lib
opencv_text454.lib
opencv_tracking454.lib
opencv_video454.lib
opencv_videoio454.lib
opencv_videostab454.lib
opencv_ximgproc454.lib
opencv_xobjdetect454.lib
opencv_xphoto454.lib

获取脚本lib.py:

import os
from os import path
#定义一个函数，方法
def scanner_file(url):
#获取当前路径下的所有文件或者文件夹
	files = os.listdir(url)
# print(files)
#拼接路径
	for f in files:
# real_path = url + "\\" + f
		real_path = path.join(url,f)
# print(real_path)
#判断是否是文件
		if path.isfile(real_path):
			# print(path.abspath(real_path))
			print(f)
			# if
#是目录(文件夹)
		elif path.isdir(real_path):
			scanner_file(real_path)
#其他情况
		else:
			print("其他情况")
			pass
scanner_file("E:\\opencv\\opencv\\cudabuild\\install\\x64\\vc15\\lib\\")#这里只是用D盘举例。

添加环境变量：

E:\opencv\opencv\cudabuild\install\x64\vc15\bin

删除之前配置opencv的bin目录，这样就配置完成。

五、代码验证

目录展示：

注意：
1、参考博客：https://blog.csdn.net/m0_59135992/article/details/125932955
2、使用Bubbliiiing的yolov7代码部署，代码地址：https://github.com/bubbliiiing/yolov7-pytorch

模型的的训练看B导即可，up主地址：https://blog.csdn.net/weixin_44791964/article/details/125827160

模型训练完成之后，在predict.py中设置mode = "export_onnx"即可生成。

注意，此处有个坑，B导的yolov7代码输出的onnx只有1class.size20*20这一层，需要在nets/yolo.py文件中修改一下。

修改之前：（在yolo.py的最下面）

 # 第三个特征层
          # y3=(batch_size, 75, 80, 80)
        ---------------------------------------------------#
        out2 = self.yolo_head_P3(P3)
        #---------------------------------------------------#
        #   第二个特征层
        #   y2=(batch_size, 75, 40, 40)
        #---------------------------------------------------#
        out1 = self.yolo_head_P4(P4)
        #------------------------------