OpenCV调用TensorFlow预训练模型
OpenCV调用TensorFlow预训练模型
【尊重原创,转载请注明出处】https://blog.csdn.net/guyuealian/article/details/80570120
强大OpenCV从自OpenCV 3.1版以来,dnn模块一直是opencv_contrib库的一部分,在3.3版中,它被提到了主仓库中。新版OpenCV dnn模块目前支持Caffe、TensorFlow、Torch、PyTorch等深度学习框架。另外,新版本中使用预训练深度学习模型的API同时兼容C++和Python
OpenCV 3.3开始就提供了读取TensoFlow模型的接口了,不过现在能支持的模型并不多。目前,我测试成功只有两个,分别是object_detection的“ssd_mobilenet_v1_coco_11_06_2017”和“ssd_inception_v2_coco_2017_11_17”预训练模型,其他模型报出各种各样的错误,反正我暂时还没有解决问题,各位神一样的网友若有新的进展,麻烦告知一声,哈哈~!
一、版本配置要求
- 1. Windows 7 和 VS2015
- 2. OpenCV3.3.1以上,本人使用的OpenCV 3.4.1,下载地址:https://github.com/opencv/opencv/releases
- 3.至于TensoFlow,鄙人只是想测试已经训练好的模型,那就可以不安装TensoFlow了,偷懒一下~!
二、模型下载和配置
1. 先用VS2015新建一个项目“opercv4tensorflow”,并在源文件目录下新建 一个“models”的文件夹,这个文件夹用于下载TensoFlow的模型。
2. 目前,只测试TensoFlow的object_detection模块的“ssd_mobilenet_v1_coco_11_06_2017”和“ssd_inception_v2_coco_2017_11_17”两个预训练模型,其他的自己看着办吧:
http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_11_06_2017.tar.gz
http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2017_11_17.tar.gz
当然了,你可以自己训练模型,也可以下载更多的预训练模型,这个可以到TensoFlow下载,下载地址是:
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
Tensorflow模型的graph结构可以保存为.pb文件或者.pbtxt文件,或者.meta文件,其中只有.pbtxt文件是可读的。在OpenCV中,每个模型.pb文件,原则上应有一个对应的文本图形定义的.pbtxt文件,当然也可能没有,在opencv_extra\testdata\dnn有些.pbtxt文件是可以对应找到,这个要看opencv会不会提供,当然,你厉害的话,可以自己按照网络定义结构写一份。
下表给出,目前我在object_detection模块中测试成功.pb文件与.pbtxt文件的对应关系,更多的模型,还请网友下方留言,支援一下,我会及时更新博客内容:
序号 | .pb文件:预训练模型 | .pbtxt文件 | 备注 |
1 | ssd_mobilenet_v1_coco_11_06_2017 | ssd_mobilenet_v1_coco.pbtxt | |
2 | ssd_inception_v2_coco_2017_11_17 | ssd_inception_v2_coco_2017_11_17.pbtxt | |
3 | |||
4 | |||
3. 上面下载的TensoFlow模型解压后,里含有重要的二进制protobuf描述的.pb文件,我们还需要对应的protobuf格式文本图形定义的.pbtxt文件,这个就需要到opencv_extra\testdata\dnn下载了
opencv_extra下载地址:https://github.com/opencv/opencv_extra/tree/master/testdata/dnn
把dnn文件夹中的“ssd_inception_v2_coco_2017_11_17.pbtxt”和“ssd_mobilenet_v1_coco.pbtxt”也下载复制到项目的models中吧
三、使用opencv读取网络模型
一切准备妥当了,下面就使用opencv C++(也可以是python的)实现读取TensoFlow训练好的网络模型。opencv dnn为我们提供TensoFlow的接口:readNetFromTensorflow,当然也有支持Caffe的接口:readNetFromCaffe。我们这里只考虑TensoFlow的模型,readNetFromTensorflow函数有多个重载函数:
/** @brief Reads a network model stored in <a href="https://www.tensorflow.org/">TensorFlow</a> framework's format. * @param model path to the .pb file with binary protobuf description of the network architecture * @param config path to the .pbtxt file that contains text graph definition in protobuf format. * Resulting Net object is built by text graph using weights from a binary one that * let us make it more flexible. * @returns Net object. */ CV_EXPORTS_W Net readNetFromTensorflow(const String &model, const String &config = String());
完整的测试代码:
#include<opencv2\opencv.hpp>#include<opencv2\dnn.hpp>#include <iostream> using namespace std;using namespace cv; const size_t inWidth = 300;const size_t inHeight = 300;const float WHRatio = inWidth / (float)inHeight;const char* classNames[] = { "background","face" };//这个需要根据训练的类别定义 int main() { Mat frame = cv::imread("2.jpg"); Size frame_size = frame.size(); String weights = "models\\ssd_inception_v2_coco_2017_11_17\\frozen_inference_graph.pb"; String prototxt = "models\\ssd_inception_v2_coco_2017_11_17.pbtxt"; dnn::Net net = cv::dnn::readNetFromTensorflow(weights, prototxt); Size cropSize; if (frame_size.width / (float)frame_size.height > WHRatio) { cropSize = Size(static_cast<int>(frame_size.height * WHRatio), frame_size.height); } else { cropSize = Size(frame_size.width, static_cast<int>(frame_size.width / WHRatio)); } Rect crop(Point((frame_size.width - cropSize.width) / 2, (frame_size.height - cropSize.height) / 2), cropSize); cv::Mat blob = cv::dnn::blobFromImage(frame, 1. / 255, Size(300, 300)); //cout << "blob size: " << blob.size << endl; net.setInput(blob); Mat output = net.forward(); //cout << "output size: " << output.size << endl; Mat detectionMat(output.size[2], output.size[3], CV_32F, output.ptr<float>()); frame = frame(crop); float confidenceThreshold = 0.20; for (int i = 0; i < detectionMat.rows; i++) { float confidence = detectionMat.at<float>(i, 2); if (confidence > confidenceThreshold) { size_t objectClass = (size_t)(detectionMat.at<float>(i, 1)); int xLeftBottom = static_cast<int>(detectionMat.at<float>(i, 3) * frame.cols); int yLeftBottom = static_cast<int>(detectionMat.at<float>(i, 4) * frame.rows); int xRightTop = static_cast<int>(detectionMat.at<float>(i, 5) * frame.cols); int yRightTop = static_cast<int>(detectionMat.at<float>(i, 6) * frame.rows); ostringstream ss; ss << confidence; String conf(ss.str()); Rect object((int)xLeftBottom, (int)yLeftBottom, (int)(xRightTop - xLeftBottom), (int)(yRightTop - yLeftBottom)); rectangle(frame, object, Scalar(0, 255, 0), 2); String label = String(classNames[objectClass]) + ": " + conf; int baseLine = 0; Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine); rectangle(frame, Rect(Point(xLeftBottom, yLeftBottom - labelSize.height), Size(labelSize.width, labelSize.height + baseLine)), Scalar(0, 255, 0), CV_FILLED); putText(frame, label, Point(xLeftBottom, yLeftBottom), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 0, 0)); } } imshow("image", frame); waitKey(0); return 0;}
运行结果:
四、常见的错误和坑
- OpenCV(3.4.1) Error: Unspecified error (FAILED: fs.is_open(). Can't open "models\ssd_inception_v2_coco_2017_1117\frozen_inference_graph.pb") in cv::dnn::ReadProtoFromBinaryFile, file C:\build\master_winpack-build-win64-vc14\opencv\modules\dnn\src\caffe\caffe_io.cpp, line 1126
出现这个问题,很大的原因是你的模型路径,错鸟!!导致不能打开
- OpenCV(3.4.1) Error: Unspecified error (Const input blob for weights not found) in cv::dnn::experimental_dnn_v4::`anonymous-namespace'::TFImporter::getConstBlob, file C:\build\master_winpack-build-win64-vc14\opencv\modules\dnn\src\tensorflow\tf_importer.cpp, line 579
出现这个问题,可能是你的.pb模型文件与.pbtxt文件不对应
五、参考资料:
https://blog.csdn.net/xingchenbingbuyu/article/details/78416887