这个例子展示了如何加载预先训练好的TensorFlow网络并使用它在C++中检测图像中的对象。

这个演示使用了一个基于[使用深度神经网络的可伸缩对象检测]的模型来检测从网络传来的图像中的人命令行。这与Android TensorFlow中使用的模型相同在相机预览中进行实时人员检测和跟踪的演示。

函数int main(int argc, char* argv[])

函数声明情况如下：

函数声明文件如下：

函数流程图如下：

函数逻辑顺序图如下：

 main.CC文件的调用关系：

函数原始代码如下：

int main(int argc, char* argv[]) {

  //这些是程序可以理解的命令行标志。它们定义了图形和输入数据的位置，以及模型期望的输入类型。如果你训练你自己的模型，或者使用其他模型，你需要更新这些模型。

  string image =

      "tensorflow/examples/multibox_detector/data/surfers.jpg";

  string graph =

      "tensorflow/examples/multibox_detector/data/"

      "multibox_model.pb";

  string box_priors =

      "tensorflow/examples/multibox_detector/data/"

      "multibox_location_priors.txt";

  int32_t input_width = 224;

  int32_t input_height = 224;

  int32_t input_mean = 128;

  int32_t input_std = 128;

  int32_t num_detections = 5;

  int32_t num_boxes = 784;

  string input_layer = "ResizeBilinear";

  string output_location_layer = "output_locations/Reshape";

  string output_score_layer = "output_scores/Reshape";

  string root_dir = "";

  string image_out = "";

  std::vector<Flag> flag_list = {

      Flag("image", &image, "image to be processed"),

      Flag("image_out", &image_out,

           "location to save output image, if desired"),

      Flag("graph", &graph, "graph to be executed"),

      Flag("box_priors", &box_priors, "name of file containing box priors"),

      Flag("input_width", &input_width, "resize image to this width in pixels"),

      Flag("input_height", &input_height,

           "resize image to this height in pixels"),

      Flag("input_mean", &input_mean, "scale pixel values to this mean"),

      Flag("input_std", &input_std, "scale pixel values to this std deviation"),

      Flag("num_detections", &num_detections,

           "number of top detections to return"),

      Flag("num_boxes", &num_boxes,

           "number of boxes defined by the location file"),

      Flag("input_layer", &input_layer, "name of input layer"),

      Flag("output_location_layer", &output_location_layer,

           "name of location output layer"),

      Flag("output_score_layer", &output_score_layer,

           "name of score output layer"),

      Flag("root_dir", &root_dir,

           "interpret image and graph file names relative to this directory"),

  };

  string usage = tensorflow::Flags::Usage(argv[0], flag_list);

  const bool parse_result = tensorflow::Flags::Parse(&argc, argv, flag_list);

  if (!parse_result) {

    LOG(ERROR) << usage;

    return -1;

  }

  //我们需要称之为TensorFlow来建立全局状态。

  tensorflow::port::InitMain(argv[0], &argc, &argv);

  if (argc > 1) {

    LOG(ERROR) << "Unknown argument " << argv[1] << "\n" << usage;

    return -1;

  }

  //首先，我们加载并初始化模型。

  std::unique_ptr<tensorflow::Session> session;

  string graph_path = tensorflow::io::JoinPath(root_dir, graph);

  Status load_graph_status = LoadGraph(graph_path, &session);

  if (!load_graph_status.ok()) {

    LOG(ERROR) << load_graph_status;

    return -1;

  }

  //以浮点数数组的形式从磁盘获取图像，并根据主图形所需的规格调整大小和规格。  std::vector<Tensor> image_tensors;

  string image_path = tensorflow::io::JoinPath(root_dir, image);

  Status read_tensor_status =

      ReadTensorFromImageFile(image_path, input_height, input_width, input_mean,

                              input_std, &image_tensors);

  if (!read_tensor_status.ok()) {

    LOG(ERROR) << read_tensor_status;

    return -1;

  }

  const Tensor& resized_tensor = image_tensors[0];

  //在模型中实际运行图像。

  std::vector<Tensor> outputs;

  Status run_status =

      session->Run({{input_layer, resized_tensor}},

                   {output_score_layer, output_location_layer}, {}, &outputs);

  if (!run_status.ok()) {

    LOG(ERROR) << "Running model failed: " << run_status;

    return -1;

  }

  Status print_status = PrintTopDetections(outputs, box_priors, num_boxes,

                                           num_detections, image_out,

                                           &image_tensors[1]);

  if (!print_status.ok()) {

    LOG(ERROR) << "Running print failed: " << print_status;

    return -1;

  }

  return 0;

}