Caffe训练好的网络对图像分类

对于训练好的Caffe 网络

输入：彩色or灰度图片

做minist 下手写识别分类，不能直接使用，需去除均值图像，同时将输入图像像素归一化到0-1直接即可。

#include <caffe/caffe.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <iosfwd>
#include <memory>
#include <string>
#include <utility>
#include <vector>

using namespace caffe;  // NOLINT(build/namespaces)
using std::string;

/* Pair (label, confidence) representing a prediction. */
/* pair(标签，置信度)  预测值 */
typedef std::pair<string, float> Prediction;
/*  分类接口类 Classifier */
class Classifier {
 public:
  Classifier(const string& model_file,
             const string& trained_file,
             const string& mean_file,
             const string& label_file);

std::vector<Prediction> Classify(const cv::Mat& img, int N = 4);   //分类，默认返回前4个预测值 数组

private:
  void SetMean(const string& mean_file);

std::vector<float> Predict(const cv::Mat& img);

void WrapInputLayer(std::vector<cv::Mat>* input_channels);

void Preprocess(const cv::Mat& img,
                  std::vector<cv::Mat>* input_channels);

private:
  shared_ptr<Net<float> > net_;            
  cv::Size input_geometry_;                 
  int num_channels_;                          //网络通道数
  cv::Mat mean_;                              //均值图像
  std::vector<string> labels_;                //目标标签数组
};

以上定义了一个分类对象类Classifier

类的实现如下：

Classifier::Classifier(const string& model_file,
                       const string& trained_file,
                       const string& mean_file,
                       const string& label_file) {
#ifdef CPU_ONLY
  Caffe::set_mode(Caffe::CPU);
#else
  Caffe::set_mode(Caffe::GPU);
#endif

/* Load the network. */
  net_.reset(new Net<float>(model_file, TEST));
  net_->CopyTrainedLayersFrom(trained_file);

CHECK_EQ(net_->num_inputs() , 1) << "Network should have exactly one input.";
  CHECK_EQ(net_->num_outputs(), 1) << "Network should have exactly one output.";

Blob<float>* input_layer = net_->input_blobs()[0];     //网络层模板 Blob
  num_channels_            = input_layer->channels();    //通道数
  CHECK(num_channels_ == 3 || num_channels_ == 1) << "Input layer should have 1 or 3 channels.";
  input_geometry_          = cv::Size(input_layer->width(), input_layer->height());

/* Load the binaryproto mean file.加载均值文件 */
  SetMean(mean_file);

/* Load labels. 加载分类标签文件*/
  std::ifstream labels(label_file.c_str());
  CHECK(labels) << "Unable to open labels file " << label_file;
  string line;
  while (std::getline(labels, line))
        labels_.push_back(string(line));

Blob<float>* output_layer = net_->output_blobs()[0];
  CHECK_EQ(labels_.size(), output_layer->channels()) << "Number of labels is different from the output layer dimension.";
}

static bool PairCompare(const std::pair<float, int>& lhs,
                        const std::pair<float, int>& rhs) {
  return lhs.first > rhs.first;
}

/* Return the indices of the top N values of vector v. */
/* 返回数组v[] 最大值的前 N 个序号数组 */
static std::vector<int> Argmax(const std::vector<float>& v, int N) {
  std::vector<std::pair<float, int> > pairs;
  for (size_t i = 0; i < v.size(); ++i)
       pairs.push_back(std::make_pair(v[i], i));

std::partial_sort(pairs.begin(), pairs.begin() + N, pairs.end(), PairCompare);

std::vector<int> result;
  for (int i = 0; i < N; ++i)
    result.push_back(pairs[i].second);
  return result;
}

/* Return the top N predictions. 分类并返回最大的前 N 个预测 */
std::vector<Prediction> Classifier::Classify(const cv::Mat& img, int N) {
  std::vector<float> output = Predict(img);

std::vector<int> maxN = Argmax(output, N);
  std::vector<Prediction> predictions;
  for (int i = 0; i < N; ++i) {
    int idx = maxN[i];
    predictions.push_back(std::make_pair(labels_[idx], output[idx])); / [(标签，置信度),...]预测值数组
  }

return predictions;
}

/* Load the mean file in binaryproto format. */
void Classifier::SetMean(const string& mean_file) {
  BlobProto blob_proto;
  ReadProtoFromBinaryFileOrDie(mean_file.c_str(), &blob_proto);

/* Convert from BlobProto to Blob<float> */
  Blob<float> mean_blob;
  mean_blob.FromProto(blob_proto);
  CHECK_EQ(mean_blob.channels(), num_channels_)
    << "Number of channels of mean file doesn't match input layer.";

/* The format of the mean file is planar 32-bit float BGR or grayscale. */
  std::vector<cv::Mat> channels;
  float* data = mean_blob.mutable_cpu_data();
  for (int i = 0; i < num_channels_; ++i) {
    /* Extract an individual channel. */
    cv::Mat channel(mean_blob.height(), mean_blob.width(), CV_32FC1, data);
    channels.push_back(channel);
    data += mean_blob.height() * mean_blob.width();
  }

/* Merge the separate channels into a single image. */
  cv::Mat mean;
  cv::merge(channels, mean);

/* Compute the global mean pixel value and create a mean image
   * filled with this value. */
  cv::Scalar channel_mean = cv::mean(mean);
  mean_ = cv::Mat(input_geometry_, mean.type(), channel_mean);
}

/*  分类 */
std::vector<float> Classifier::Predict(const cv::Mat& img) {
  Blob<float>* input_layer = net_->input_blobs()[0];
  input_layer->Reshape(1, num_channels_,
                       input_geometry_.height, input_geometry_.width);
  /* Forward dimension change to all layers. */
  net_->Reshape();
  std::vector<cv::Mat> input_channels;
  WrapInputLayer(&input_channels);

Preprocess(img, &input_channels);       //数据预处理

net_->ForwardPrefilled();

/* Copy the output layer to a std::vector */
  Blob<float>* output_layer = net_->output_blobs()[0];
  const float* begin = output_layer->cpu_data();
  const float* end = begin + output_layer->channels();
  return std::vector<float>(begin, end);
}

/* Wrap the input layer of the network in separate cv::Mat objects
 * (one per channel). This way we save one memcpy operation and we
 * don't need to rely on cudaMemcpy2D. The last preprocessing
 * operation will write the separate channels directly to the input
 * layer.
  */
void Classifier::WrapInputLayer(std::vector<cv::Mat>* input_channels) {

Blob<float>* input_layer = net_->input_blobs()[0];

int width  = input_layer->width();
  int height = input_layer->height();
  float* input_data = input_layer->mutable_cpu_data();
  for (int i = 0; i < input_layer->channels(); ++i) {
    cv::Mat channel(height, width, CV_32FC1, input_data);
    input_channels->push_back(channel);
    input_data += width * height;
  }
}
//数据预处理
void Classifier::Preprocess(const cv::Mat& img,
                            std::vector<cv::Mat>* input_channels) {
  /* Convert the input image to the input image format of the network. */
  cv::Mat sample;
  //通道数据根据设置进行转换
  if (img.channels() == 3 && num_channels_ == 1)
    cv::cvtColor(img, sample, CV_BGR2GRAY);
  else if (img.channels() == 4 && num_channels_ == 1)
    cv::cvtColor(img, sample, CV_BGRA2GRAY);
  else if (img.channels() == 4 && num_channels_ == 3)
    cv::cvtColor(img, sample, CV_BGRA2BGR);
  else if (img.channels() == 1 && num_channels_ == 3)
    cv::cvtColor(img, sample, CV_GRAY2BGR);
  else
    sample = img;

cv::Mat sample_resized;
  if (sample.size() != input_geometry_)
    cv::resize(sample, sample_resized, input_geometry_);
  else
    sample_resized = sample;

cv::Mat sample_float;
  if (num_channels_ == 3)
    sample_resized.convertTo(sample_float, CV_32FC3);     // 三通道(彩色)
  else
    sample_resized.convertTo(sample_float, CV_32FC1);     // 单通道    (灰度)

cv::Mat sample_normalized;
  cv::subtract(sample_float, mean_, sample_normalized);

/* This operation will write the separate BGR planes directly to the
   * input layer of the network because it is wrapped by the cv::Mat
   * objects in input_channels.
   此操作将数据 BGR 直接写入输入层对象input_channels */
  cv::split(sample_normalized, *input_channels);

CHECK(reinterpret_cast<float*>(input_channels->at(0).data)
        == net_->input_blobs()[0]->cpu_data())
    << "Input channels are not wrapping the input layer of the network.";
}

对以上代码做了一些简单的注释，需要说明的是分类后的返回结果默认置信度最大的前5个类型，

对于分类对象的调用如下：

//==============================================================
//         main()
//==============================================================

int main(int argc, char** argv) {
  if (argc != 6) {
    std::cerr << "Usage: " << argv[0]
              << " deploy.prototxt  network.caffemodel"
              << " mean.binaryproto labels.txt img.jpg" << std::endl;
    return 1;
  }

::google::InitGoogleLogging(argv[0]);

string model_file   = argv[1];
  string trained_file = argv[2];
  string mean_file    = argv[3];
  string label_file   = argv[4];
  Classifier classifier(model_file, trained_file, mean_file, label_file); //创建分类器

string file = argv[5];

std::cout << "---------- Prediction for "<< file << " ----------" << std::endl;

cv::Mat img = cv::imread(file, -1);          //读取待分类图像
  CHECK(!img.empty()) << "Unable to decode image " << file;
  std::vector<Prediction> predictions = classifier.Classify(img);    //分类

/* Print the top N predictions. 打印前N 个预测值*/
  for (size_t i = 0; i < predictions.size(); ++i) {
    Prediction p = predictions[i];
    std::cout << std::fixed << std::setprecision(4) << p.second << " - \""
              << p.first << "\"" << std::endl;
  }
}