diff --git a/deploy/cpp/demo/classifier.cpp b/deploy/cpp/demo/classifier.cpp
index a4b3bfa3c5695f663a49862d49da9848ae86c4b0..57c4f09c3d8fe4c1dd48c6afd350cca581a2d194 100644
--- a/deploy/cpp/demo/classifier.cpp
+++ b/deploy/cpp/demo/classifier.cpp
@@ -34,6 +34,7 @@ DEFINE_string(key, "", "key of encryption");
 DEFINE_string(image, "", "Path of test image file");
 DEFINE_string(image_list, "", "Path of test image list file");
 DEFINE_int32(batch_size, 1, "Batch size of infering");
+DEFINE_int32(thread_num, omp_get_num_procs(), "Number of preprocessing threads");
 
 int main(int argc, char** argv) {
   // Parsing command-line
@@ -75,12 +76,13 @@ int main(int argc, char** argv) {
       int im_vec_size = std::min((int)image_paths.size(), i + FLAGS_batch_size);      
       std::vector<cv::Mat> im_vec(im_vec_size - i);
       std::vector<PaddleX::ClsResult> results(im_vec_size - i, PaddleX::ClsResult());
-      #pragma omp parallel for num_threads(im_vec_size - i)
+      int thread_num = std::min(FLAGS_thread_num, im_vec_size - i);
+      #pragma omp parallel for num_threads(thread_num)
       for(int j = i; j < im_vec_size; ++j){
         im_vec[j - i] = std::move(cv::imread(image_paths[j], 1));
       }
       auto imread_end = system_clock::now();
-      model.predict(im_vec, results);
+      model.predict(im_vec, results, thread_num);
 
       auto imread_duration = duration_cast<microseconds>(imread_end - start);
       total_imread_time_s += double(imread_duration.count()) * microseconds::period::num / microseconds::period::den;
diff --git a/deploy/cpp/demo/detector.cpp b/deploy/cpp/demo/detector.cpp
index 5068a35ad49791564f625fa140a2a156ecf58ed8..956a7181808defbe7df127ea06d298f3757dce4b 100644
--- a/deploy/cpp/demo/detector.cpp
+++ b/deploy/cpp/demo/detector.cpp
@@ -36,12 +36,14 @@ DEFINE_string(key, "", "key of encryption");
 DEFINE_string(image, "", "Path of test image file");
 DEFINE_string(image_list, "", "Path of test image list file");
 DEFINE_string(save_dir, "output", "Path to save visualized image");
-DEFINE_int32(batch_size, 1, "");
+DEFINE_int32(batch_size, 1, "Batch size of infering");
+DEFINE_double(threshold, 0.5, "The minimum scores of target boxes which are shown");
+DEFINE_int32(thread_num, omp_get_num_procs(), "Number of preprocessing threads");
 
 int main(int argc, char** argv) {
   // 解析命令行参数
   google::ParseCommandLineFlags(&argc, &argv, true);
-
+  
   if (FLAGS_model_dir == "") {
     std::cerr << "--model_dir need to be defined" << std::endl;
     return -1;
@@ -50,7 +52,7 @@ int main(int argc, char** argv) {
     std::cerr << "--image or --image_list need to be defined" << std::endl;
     return -1;
   }
-
+  std::cout << "Thread num: " << FLAGS_thread_num << std::endl;
   // 加载模型
   PaddleX::Model model;
   model.Init(FLAGS_model_dir, FLAGS_use_gpu, FLAGS_use_trt, FLAGS_gpu_id, FLAGS_key, FLAGS_batch_size);
@@ -78,12 +80,13 @@ int main(int argc, char** argv) {
       int im_vec_size = std::min((int)image_paths.size(), i + FLAGS_batch_size);
       std::vector<cv::Mat> im_vec(im_vec_size - i);
       std::vector<PaddleX::DetResult> results(im_vec_size - i, PaddleX::DetResult());
-      #pragma omp parallel for num_threads(im_vec_size - i)
+      int thread_num = std::min(FLAGS_thread_num, im_vec_size - i);
+      #pragma omp parallel for num_threads(thread_num)
       for(int j = i; j < im_vec_size; ++j){
         im_vec[j - i] = std::move(cv::imread(image_paths[j], 1));
       }
       auto imread_end = system_clock::now();
-      model.predict(im_vec, results);
+      model.predict(im_vec, results, thread_num);
       auto imread_duration = duration_cast<microseconds>(imread_end - start);
       total_imread_time_s += double(imread_duration.count()) * microseconds::period::num / microseconds::period::den;
       auto end = system_clock::now();
@@ -106,7 +109,7 @@ int main(int argc, char** argv) {
       // 可视化
       for(int j = 0; j < im_vec_size - i; ++j) {
         cv::Mat vis_img =
-            PaddleX::Visualize(im_vec[j], results[j], model.labels, colormap, 0.5);
+            PaddleX::Visualize(im_vec[j], results[j], model.labels, colormap, FLAGS_threshold);
         std::string save_path =
             PaddleX::generate_save_path(FLAGS_save_dir, image_paths[i + j]);
         cv::imwrite(save_path, vis_img);
@@ -130,7 +133,7 @@ int main(int argc, char** argv) {
 
     // 可视化
     cv::Mat vis_img =
-        PaddleX::Visualize(im, result, model.labels, colormap, 0.5);
+        PaddleX::Visualize(im, result, model.labels, colormap, FLAGS_threshold);
     std::string save_path =
         PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_image);
     cv::imwrite(save_path, vis_img);
diff --git a/deploy/cpp/demo/segmenter.cpp b/deploy/cpp/demo/segmenter.cpp
index 23bdd1f955e089a477d9bc029b292a59d9434475..87e1afdd41c3116b2027c4846d08dcfbb9a7dff6 100644
--- a/deploy/cpp/demo/segmenter.cpp
+++ b/deploy/cpp/demo/segmenter.cpp
@@ -36,6 +36,7 @@ DEFINE_string(image, "", "Path of test image file");
 DEFINE_string(image_list, "", "Path of test image list file");
 DEFINE_string(save_dir, "output", "Path to save visualized image");
 DEFINE_int32(batch_size, 1, "Batch size of infering");
+DEFINE_int32(thread_num, omp_get_num_procs(), "Number of preprocessing threads");
 
 int main(int argc, char** argv) {
   // 解析命令行参数
@@ -76,12 +77,13 @@ int main(int argc, char** argv) {
       int im_vec_size = std::min((int)image_paths.size(), i + FLAGS_batch_size);
       std::vector<cv::Mat> im_vec(im_vec_size - i);
       std::vector<PaddleX::SegResult> results(im_vec_size - i, PaddleX::SegResult());
-      #pragma omp parallel for num_threads(im_vec_size - i)
+      int thread_num = std::min(FLAGS_thread_num, im_vec_size - i);
+      #pragma omp parallel for num_threads(thread_num)
       for(int j = i; j < im_vec_size; ++j){
         im_vec[j - i] = std::move(cv::imread(image_paths[j], 1));
       }
       auto imread_end = system_clock::now();
-      model.predict(im_vec, results);
+      model.predict(im_vec, results, thread_num);
       auto imread_duration = duration_cast<microseconds>(imread_end - start);
       total_imread_time_s += double(imread_duration.count()) * microseconds::period::num / microseconds::period::den;
       auto end = system_clock::now();
diff --git a/deploy/cpp/include/paddlex/paddlex.h b/deploy/cpp/include/paddlex/paddlex.h
index bcddd87845f8edff135047a0edbf9706c8579316..c0a48768b334d3a62b22f201989706a393a70a92 100644
--- a/deploy/cpp/include/paddlex/paddlex.h
+++ b/deploy/cpp/include/paddlex/paddlex.h
@@ -61,19 +61,19 @@ class Model {
 
   bool preprocess(const cv::Mat& input_im, ImageBlob* blob);
   
-  bool preprocess(const std::vector<cv::Mat> &input_im_batch, std::vector<ImageBlob> &blob_batch);
+  bool preprocess(const std::vector<cv::Mat> &input_im_batch, std::vector<ImageBlob> &blob_batch, int thread_num = 1);
 
   bool predict(const cv::Mat& im, ClsResult* result);
 
-  bool predict(const std::vector<cv::Mat> &im_batch, std::vector<ClsResult> &results);
+  bool predict(const std::vector<cv::Mat> &im_batch, std::vector<ClsResult> &results, int thread_num = 1);
 
   bool predict(const cv::Mat& im, DetResult* result);
 
-  bool predict(const std::vector<cv::Mat> &im_batch, std::vector<DetResult> &result);
+  bool predict(const std::vector<cv::Mat> &im_batch, std::vector<DetResult> &result, int thread_num = 1);
   
   bool predict(const cv::Mat& im, SegResult* result);
 
-  bool predict(const std::vector<cv::Mat> &im_batch, std::vector<SegResult> &result);
+  bool predict(const std::vector<cv::Mat> &im_batch, std::vector<SegResult> &result, int thread_num = 1);
   
   std::string type;
   std::string name;
diff --git a/deploy/cpp/src/paddlex.cpp b/deploy/cpp/src/paddlex.cpp
index 5dd1237cc7280fcc6fe6dd010463ae1facb57c2a..f52cf7afe36790e7a83d8e6539a23cf8e9eba1f5 100644
--- a/deploy/cpp/src/paddlex.cpp
+++ b/deploy/cpp/src/paddlex.cpp
@@ -110,12 +110,13 @@ bool Model::preprocess(const cv::Mat& input_im, ImageBlob* blob) {
 }
 
 // use openmp
-bool Model::preprocess(const std::vector<cv::Mat> &input_im_batch, std::vector<ImageBlob> &blob_batch) {
+bool Model::preprocess(const std::vector<cv::Mat> &input_im_batch, std::vector<ImageBlob> &blob_batch, int thread_num) {
   int batch_size = input_im_batch.size();
   bool success = true;
   int max_h = -1;
   int max_w = -1;
-  #pragma omp parallel for num_threads(batch_size)
+  thread_num = std::min(thread_num, batch_size);
+  #pragma omp parallel for num_threads(thread_num)
   for(int i = 0; i < input_im_batch.size(); ++i) {
     cv::Mat im = input_im_batch[i].clone();
     if(!transforms_.Run(&im, &blob_batch[i])){
@@ -164,7 +165,7 @@ bool Model::predict(const cv::Mat& im, ClsResult* result) {
   return true;
 }
 
-bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<ClsResult> &results) {
+bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<ClsResult> &results, int thread_num) {
   for(auto &inputs: inputs_batch_) {
     inputs.clear();
   }
@@ -180,7 +181,7 @@ bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<ClsResult>
     return false;
   }
   // 处理输入图像
-  if (!preprocess(im_batch, inputs_batch_)) {
+  if (!preprocess(im_batch, inputs_batch_, thread_num)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
   }
@@ -326,7 +327,7 @@ bool Model::predict(const cv::Mat& im, DetResult* result) {
   return true;
 }
 
-bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<DetResult> &result) {
+bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<DetResult> &result, int thread_num) {
   if (type == "classifier") {
     std::cerr << "Loading model is a 'classifier', ClsResult should be passed "
                  "to function predict()!"
@@ -341,7 +342,7 @@ bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<DetResult>
 
   int batch_size = im_batch.size();
   // 处理输入图像
-  if (!preprocess(im_batch, inputs_batch_)) {
+  if (!preprocess(im_batch, inputs_batch_, thread_num)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
   }
@@ -357,7 +358,8 @@ bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<DetResult>
                   << ", " << inputs_batch_[i].new_im_size_[1] 
                   <<  ")" << std::endl;
       }
-      #pragma omp parallel for num_threads(batch_size)
+      thread_num = std::min(thread_num, batch_size);
+      #pragma omp parallel for num_threads(thread_num)
       for(int i = 0; i < batch_size; ++i) {
         int h = inputs_batch_[i].new_im_size_[0];
         int w = inputs_batch_[i].new_im_size_[1];
@@ -597,7 +599,7 @@ bool Model::predict(const cv::Mat& im, SegResult* result) {
   return true;
 }
 
-bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<SegResult> &result) {
+bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<SegResult> &result, int thread_num) {
   for(auto &inputs: inputs_batch_) {
     inputs.clear();
   }
@@ -614,7 +616,7 @@ bool Model::predict(const std::vector<cv::Mat> &im_batch, std::vector<SegResult>
   }
 
   // 处理输入图像
-  if (!preprocess(im_batch, inputs_batch_)) {
+  if (!preprocess(im_batch, inputs_batch_, thread_num)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
   }