diff --git a/deploy/cpp/README.md b/deploy/cpp/README.md
index c9d1c08de5cbea3c09980ff138c2ff7a417e93bd..790a3a6ae5a28a40416ac0cd940d7760b1b1c509 100644
--- a/deploy/cpp/README.md
+++ b/deploy/cpp/README.md
@@ -25,7 +25,7 @@
 ## 2.主要目录和文件
 
 ```
-inference
+cpp
 ├── demo.cpp # 演示加载模型、读入数据、完成预测任务C++代码
 |
 ├── conf
@@ -90,6 +90,8 @@ deeplabv3p_xception65_humanseg
 DEPLOY:
     # 是否使用GPU预测
     USE_GPU: 1
+    # 是否是PaddleSeg 0.3.0新版本模型
+    USE_PR : 1
     # 模型和参数文件所在目录路径
     MODEL_PATH: "/root/projects/models/deeplabv3p_xception65_humanseg"
     # 模型文件名
@@ -125,11 +127,11 @@ DEPLOY:
 
 `Linux` 系统中执行以下命令：
 ```shell
-./demo --conf=/root/projects/PaddleSeg/inference/conf/humanseg.yaml --input_dir=/root/projects/PaddleSeg/inference/images/humanseg/
+./demo --conf=/root/projects/PaddleSeg/deploy/cpp/conf/humanseg.yaml --input_dir=/root/projects/PaddleSeg/deploy/cpp/images/humanseg/
 ```
 `Windows` 中执行以下命令:
 ```shell
-D:\projects\PaddleSeg\inference\build\Release>demo.exe --conf=D:\\projects\\PaddleSeg\\inference\\conf\\humanseg.yaml --input_dir=D:\\projects\\PaddleSeg\\inference\\images\humanseg\\
+D:\projects\PaddleSeg\deploy\cpp\build\Release>demo.exe --conf=D:\\projects\\PaddleSeg\\deploy\\cpp\\conf\\humanseg.yaml --input_dir=D:\\projects\\PaddleSeg\\deploy\\cpp\\images\humanseg\\
 ```
 
 
@@ -141,7 +143,7 @@ D:\projects\PaddleSeg\inference\build\Release>demo.exe --conf=D:\\projects\\Padd
 | input_dir | 需要预测的图片目录 |
 
 
-配置文件说明请参考上一步，样例程序会扫描input_dir目录下的所有以**jpg或jpeg**为后缀的图片，并生成对应的预测结果（若input_dir目录下没有以**jpg或jpeg**为后缀的图片，程序会报错）。图像分割会对`demo.jpg`的每个像素进行分类，其预测的结果保存在`demo_jpg.png`中。分割预测结果的图不能直接看到效果，必须经过可视化处理。对于二分类的图像分割模型，样例程序自动将预测结果转换成可视化结果，保存在`demo_jpg_scoremap.png`中， 原始尺寸的预测结果在`demo_jpg_recover.png`中，如下图。对于**多分类**的图像分割模型，请参考[可视化脚本使用方法](./docs/vis.md)。
+配置文件说明请参考上一步，样例程序会扫描input_dir目录下的所有以**jpg或jpeg**为后缀的图片，并生成对应的预测结果（若input_dir目录下没有以**jpg或jpeg**为后缀的图片，程序会报错）。图像分割会对`demo.jpg`的每个像素进行分类，其预测的结果保存在`demo_jpg_mask.png`中。分割预测结果的图不能直接看到效果，必须经过可视化处理。对于二分类的图像分割模型。如果需要对预测结果进行**可视化**，请参考[可视化脚本使用方法](./docs/vis.md)。
 
 输入原图  
 ![avatar](images/humanseg/demo2.jpeg)
diff --git a/deploy/cpp/demo.cpp b/deploy/cpp/demo.cpp
index 2202b31b739bf18682fdf468b36ffe4e9e434726..729b36dde098b4ddaa6d0ed1d0da320c82e1a304 100644
--- a/deploy/cpp/demo.cpp
+++ b/deploy/cpp/demo.cpp
@@ -24,7 +24,7 @@ int main(int argc, char** argv) {
     google::ParseCommandLineFlags(&argc, &argv, true);
     if (FLAGS_conf.empty() || FLAGS_input_dir.empty()) {
         std::cout << "Usage: ./predictor --conf=/config/path/to/your/model "
-                  << "--input_dir=/directory/of/your/input/images";
+                  << "--input_dir=/directory/of/your/input/images" << std::endl;
         return -1;
     }
     // 1. create a predictor and init it with conf
diff --git a/deploy/cpp/docs/linux_build.md b/deploy/cpp/docs/linux_build.md
index 75a16bbea7499eaa007884b0b1f16126eacca56a..b294eb698e75df29f2607f0e73b27075c7b8d397 100644
--- a/deploy/cpp/docs/linux_build.md
+++ b/deploy/cpp/docs/linux_build.md
@@ -16,7 +16,7 @@
 1. `mkdir -p /root/projects/ && cd /root/projects`
 2. `git clone https://github.com/PaddlePaddle/PaddleSeg.git`
 
-`C++`预测代码在`/root/projects/PaddleSeg/inference` 目录，该目录不依赖任何`PaddleSeg`下其他目录。
+`C++`预测代码在`/root/projects/PaddleSeg/deploy/cpp` 目录，该目录不依赖任何`PaddleSeg`下其他目录。
 
 
 ### Step2: 下载PaddlePaddle C++ 预测库 fluid_inference
@@ -25,9 +25,9 @@ PaddlePaddle C++ 预测库主要分为CPU版本和GPU版本。其中，针对不
 
 |  版本   | 链接  |
 |  ----  | ----  |
-| CPU版本  | [fluid_inference.tgz](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_linux_cpu_1.6.1.tgz) |
-| CUDA 9.0版本  | [fluid_inference.tgz](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_linux_cuda97_1.6.1.tgz) |
-| CUDA 10.0版本  | [fluid_inference.tgz](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_linux_cuda10_1.6.1.tgz) |
+| CPU版本  | [fluid_inference.tgz](https://paddle-inference-lib.bj.bcebos.com/1.6.1-cpu-avx-mkl/fluid_inference.tgz) |
+| CUDA 9.0版本  | [fluid_inference.tgz](https://paddle-inference-lib.bj.bcebos.com/1.6.1-gpu-cuda9-cudnn7-avx-mkl/fluid_inference.tgz) |
+| CUDA 10.0版本  | [fluid_inference.tgz](https://paddle-inference-lib.bj.bcebos.com/1.6.1-gpu-cuda10-cudnn7-avx-mkl/fluid_inference.tgz) |
 
 
 针对不同的CPU类型、不同的指令集，官方提供更多可用的预测库版本，目前已经推出1.6版本的预测库。其余版本具体请参考以下链接:[C++预测库下载列表](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_usage/deploy/inference/build_and_install_lib_cn.html)
@@ -75,7 +75,7 @@ make install
 在使用**GPU版本**预测库进行编译时，可执行下列操作。**注意**把对应的参数改为你的上述依赖库实际路径：
 
 ```shell
-cd /root/projects/PaddleSeg/inference
+cd /root/projects/PaddleSeg/deploy/cpp
 mkdir build && cd build
 cmake .. -DWITH_GPU=ON  -DPADDLE_DIR=/root/projects/fluid_inference -DCUDA_LIB=/usr/local/cuda/lib64/ -DOPENCV_DIR=/root/projects/opencv3/ -DCUDNN_LIB=/usr/local/cuda/lib64/ -DWITH_STATIC_LIB=OFF
 make
@@ -83,7 +83,7 @@ make
 
 在使用**CPU版本**预测库进行编译时，可执行下列操作。
 ```shell
-cd /root/projects/PaddleSeg/inference
+cd /root/projects/PaddleSeg/cpp
 
 mkdir build && cd build
 cmake .. -DWITH_GPU=OFF  -DPADDLE_DIR=/root/projects/fluid_inference -DOPENCV_DIR=/root/projects/opencv3/ -DWITH_STATIC_LIB=OFF
@@ -98,4 +98,4 @@ make
 ./demo --conf=/path/to/your/conf --input_dir=/path/to/your/input/data/directory
 ```
 
-更详细说明请参考README文档： [预测和可视化部分](../README.md)
\ No newline at end of file
+更详细说明请参考README文档： [预测和可视化部分](../README.md)
diff --git a/deploy/cpp/docs/vis.md b/deploy/cpp/docs/vis.md
index bed5c7adba7c9739fa0a10a46d04ff6b060eeeb2..60a588f5f92174c7ffba5706fc62d36a2281c0cb 100644
--- a/deploy/cpp/docs/vis.md
+++ b/deploy/cpp/docs/vis.md
@@ -12,7 +12,7 @@ cd inference/tools/
 # 拷贝保存分割预测结果的图片到本目录
 cp XXX/demo_jpg.png .
 # 运行可视化脚本
-python visualize.py demo.jpg demo_jpg.png vis_result.png
+python visualize.py demo.jpg demo_jpg_mask.png vis_result.png
 ```
 
 以下为上述运行可视化脚本例子中每个参数的含义，请根据测试机器中图片的**实际路径**修改对应参数。
diff --git a/deploy/cpp/docs/windows_vs2015_build.md b/deploy/cpp/docs/windows_vs2015_build.md
index f0c96a18b1204b434653be8cd29dce57d229d10c..1c975015d3967d341051b9c6d5c33abb17aed7b1 100644
--- a/deploy/cpp/docs/windows_vs2015_build.md
+++ b/deploy/cpp/docs/windows_vs2015_build.md
@@ -15,7 +15,7 @@
 1. 打开`cmd`, 执行 `cd /d D:\projects`
 2. `git clone http://gitlab.baidu.com/Paddle/PaddleSeg.git`
 
-`C++`预测库代码在`D:\projects\PaddleSeg\inference` 目录，该目录不依赖任何`PaddleSeg`下其他目录。
+`C++`预测库代码在`D:\projects\PaddleSeg\deploy\cpp` 目录，该目录不依赖任何`PaddleSeg`下其他目录。
 
 
 ### Step2: 下载PaddlePaddle C++ 预测库 fluid_inference
@@ -24,9 +24,9 @@ PaddlePaddle C++ 预测库主要分为两大版本：CPU版本和GPU版本。其
 
 |  版本   | 链接  |
 |  ----  | ----  |
-| CPU版本  | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_install_dir_win_cpu_1.6.zip) |
-| CUDA 9.0版本  | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda9_1.6.1.zip) |
-| CUDA 10.0版本  | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda10_1.6.1.zip) |
+| CPU版本  | [fluid_inference_install_dir.zip](https://paddle-wheel.bj.bcebos.com/1.6.2/win-infer/mkl/cpu/fluid_inference_install_dir.zip) |
+| CUDA 9.0版本  | [fluid_inference_install_dir.zip](https://paddle-wheel.bj.bcebos.com/1.6.2/win-infer/mkl/post97/fluid_inference_install_dir.zip) |
+| CUDA 10.0版本  | [fluid_inference_install_dir.zip](https://paddle-wheel.bj.bcebos.com/1.6.2/win-infer/mkl/post107/fluid_inference_install_dir.zip) |
 
 解压后`D:\projects\fluid_inference`目录包含内容为：
 ```
@@ -70,19 +70,19 @@ call "C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\vcvarsall.bat" amd6
 
 ```bash
 # 切换到预测库所在目录
-cd /d D:\projects\PaddleSeg\inference\
+cd /d D:\projects\PaddleSeg\deply\cpp\
 # 创建构建目录, 重新构建只需要删除该目录即可
 mkdir build
 cd build
 # cmake构建VS项目
-D:\projects\PaddleSeg\inference\build> cmake .. -G "Visual Studio 14 2015 Win64" -DWITH_GPU=ON -DPADDLE_DIR=D:\projects\fluid_inference -DCUDA_LIB=D:\projects\cudalib\v9.0\lib\x64 -DOPENCV_DIR=D:\projects\opencv -T host=x64
+D:\projects\PaddleSeg\deploy\cpp\build> cmake .. -G "Visual Studio 14 2015 Win64" -DWITH_GPU=ON -DPADDLE_DIR=D:\projects\fluid_inference -DCUDA_LIB=D:\projects\cudalib\v9.0\lib\x64 -DOPENCV_DIR=D:\projects\opencv -T host=x64
 ```
 
 在使用**CPU版本**预测库进行编译时，可执行下列操作。
 
 ```bash
 # 切换到预测库所在目录
-cd /d D:\projects\PaddleSeg\inference\
+cd /d D:\projects\PaddleSeg\deploy\cpp\
 # 创建构建目录, 重新构建只需要删除该目录即可
 mkdir build
 cd build
@@ -102,7 +102,7 @@ D:\projects\PaddleSeg\inference\build> msbuild /m /p:Configuration=Release cpp_i
 
 上述`Visual Studio 2015`编译产出的可执行文件在`build\release`目录下，切换到该目录：
 ```
-cd /d D:\projects\PaddleSeg\inference\build\release
+cd /d D:\projects\PaddleSeg\deploy\cpp\build\release
 ```
 
 之后执行命令：
diff --git a/deploy/cpp/docs/windows_vs2019_build.md b/deploy/cpp/docs/windows_vs2019_build.md
index 890844674db848177c4f859f4a8b8ef8d7360fa7..5862740b19c4d146ce38b458cd8cbe76f7a84747 100644
--- a/deploy/cpp/docs/windows_vs2019_build.md
+++ b/deploy/cpp/docs/windows_vs2019_build.md
@@ -15,7 +15,7 @@ Windows 平台下，我们使用`Visual Studio 2015` 和 `Visual Studio 2019 Com
 
 ### Step1: 下载代码
 
-1. 点击下载源代码：[下载地址](https://github.com/PaddlePaddle/PaddleSeg/archive/release/v0.2.0.zip)
+1. 点击下载源代码：[下载地址](https://github.com/PaddlePaddle/PaddleSeg/archive/release/v0.3.0.zip)
 2. 解压，解压后目录重命名为`PaddleSeg`
 
 以下代码目录路径为`D:\projects\PaddleSeg` 为例。
@@ -27,9 +27,9 @@ PaddlePaddle C++ 预测库主要分为两大版本：CPU版本和GPU版本。其
 
 |  版本   | 链接  |
 |  ----  | ----  |
-| CPU版本  | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_install_dir_win_cpu_1.6.zip) |
-| CUDA 9.0版本  | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda9_1.6.1.zip) |
-| CUDA 10.0版本  | [fluid_inference_install_dir.zip](https://bj.bcebos.com/paddlehub/paddle_inference_lib/fluid_inference_install_dir_win_cuda10_1.6.1.zip) |
+| CPU版本  | [fluid_inference_install_dir.zip](https://paddle-wheel.bj.bcebos.com/1.6.1/win-infer/mkl/cpu/fluid_inference_install_dir.zip) |
+| CUDA 9.0版本  | [fluid_inference_install_dir.zip](https://paddle-wheel.bj.bcebos.com/1.6.1/win-infer/mkl/post97/fluid_inference_install_dir.zip) |
+| CUDA 10.0版本  | [fluid_inference_install_dir.zip](https://paddle-wheel.bj.bcebos.com/1.6.1/win-infer/mkl/post107/fluid_inference_install_dir.zip) |
 
 解压后`D:\projects\fluid_inference`目录包含内容为：
 ```
@@ -74,6 +74,7 @@ fluid_inference
 | *CUDA_LIB  | CUDA的库路径 |
 | OPENCV_DIR  | OpenCV的安装路径 |
 | PADDLE_DIR | Paddle预测库的路径 |
+
 **注意**在使用CPU版本预测库时，需要把CUDA_LIB的勾去掉。
 ![step4](https://paddleseg.bj.bcebos.com/inference/vs2019_step5.png)
 
@@ -89,7 +90,7 @@ fluid_inference
 上述`Visual Studio 2019`编译产出的可执行文件在`out\build\x64-Release`目录下，打开`cmd`，并切换到该目录：
 
 ```
-cd /d D:\projects\PaddleSeg\inference\out\build\x64-Release
+cd /d D:\projects\PaddleSeg\deploy\cpp\out\build\x64-Release
 ```
 
 之后执行命令：
diff --git a/deploy/cpp/predictor/seg_predictor.cpp b/deploy/cpp/predictor/seg_predictor.cpp
index 788748ff6de2abb1aba89615955334f45ce23205..2580148c52115edfa60219e4f5cf7f572c865d23 100644
--- a/deploy/cpp/predictor/seg_predictor.cpp
+++ b/deploy/cpp/predictor/seg_predictor.cpp
@@ -83,7 +83,6 @@ namespace PaddleSolution {
 
         int blob_out_len = length;
         int seg_out_len = eval_height * eval_width * eval_num_class;
-
         if (blob_out_len != seg_out_len) {
             LOG(ERROR) << " [FATAL] unequal: input vs output [" <<
                 seg_out_len << "|" << blob_out_len << "]" << std::endl;
@@ -99,23 +98,20 @@ namespace PaddleSolution {
         std::string nname(fname);
         auto pos = fname.rfind(".");
         nname[pos] = '_';
-        std::string mask_save_name = nname + ".png";
+        std::string mask_save_name = nname + "_mask.png";
         cv::imwrite(mask_save_name, mask_png);
         cv::Mat scoremap_png = cv::Mat(eval_height, eval_width, CV_8UC1);
         scoremap_png.data = _scoremap.data();
-        std::string scoremap_save_name = nname
-                                       + std::string("_scoremap.png");
+        std::string scoremap_save_name = nname + std::string("_scoremap.png");
         cv::imwrite(scoremap_save_name, scoremap_png);
         std::cout << "save mask of [" << fname << "] done" << std::endl;
 
         if (height && width) {
             int recover_height = *height;
             int recover_width = *width;
-            cv::Mat recover_png = cv::Mat(recover_height,
-                                          recover_width, CV_8UC1);
+            cv::Mat recover_png = cv::Mat(recover_height, recover_width, CV_8UC1);
             cv::resize(scoremap_png, recover_png,
-                       cv::Size(recover_width, recover_height),
-                       0, 0, cv::INTER_CUBIC);
+                cv::Size(recover_width, recover_height), 0, 0, cv::INTER_CUBIC);
             std::string recover_name = nname + std::string("_recover.png");
             cv::imwrite(recover_name, recover_png);
         }
@@ -176,8 +172,13 @@ namespace PaddleSolution {
             }
             paddle::PaddleTensor im_tensor;
             im_tensor.name = "image";
-            im_tensor.shape = std::vector<int>{ batch_size, channels,
-                                                 eval_height, eval_width };
+            if (!_model_config._use_pr) {
+                im_tensor.shape = std::vector<int>{ batch_size, channels,
+                                                eval_height, eval_width };
+            } else {
+                im_tensor.shape = std::vector<int>{ batch_size, eval_height,
+                                                eval_width, channels};
+            }
             im_tensor.data.Reset(input_buffer.data(),
                                  real_buffer_size * sizeof(float));
             im_tensor.dtype = paddle::PaddleDType::FLOAT32;
@@ -202,19 +203,45 @@ namespace PaddleSolution {
                 std::cout << _outputs[0].shape[j] << ",";
             }
             std::cout << ")" << std::endl;
-            const size_t nums = _outputs.front().data.length()
-                              / sizeof(float);
-            if (out_num % batch_size != 0 || out_num != nums) {
-                LOG(ERROR) << "outputs data size mismatch with shape size.";
+
+            size_t nums = _outputs.front().data.length() / sizeof(float);
+            if (_model_config._use_pr) {
+                nums = _outputs.front().data.length() / sizeof(int64_t);
+            }
+            // size mismatch checking
+            bool size_mismatch = out_num % batch_size;
+            size_mismatch |= (!_model_config._use_pr) && (nums != out_num);
+            size_mismatch |= _model_config._use_pr && (nums != eval_height * eval_width);
+            if (size_mismatch) {
+                LOG(ERROR) << "output with a unexpected size";
                 return -1;
             }
 
+            if (_model_config._use_pr) {
+                std::vector<uchar> out_data;
+                out_data.resize(out_num);
+                auto addr = reinterpret_cast<int64_t*>(_outputs[0].data.data());
+                for (int r = 0; r < out_num; ++r) {
+                    out_data[r] = (int)(addr[r]);
+                }
+                for (int r = 0; r < batch_size; ++r) {
+                    cv::Mat mask_png = cv::Mat(eval_height, eval_width, CV_8UC1);
+                    mask_png.data = out_data.data() + eval_height*eval_width*r;
+                    auto name = imgs_batch[r];
+                    auto pos = name.rfind(".");
+                    name[pos] = '_';
+                    std::string mask_save_name = name + "_mask.png";
+                    cv::imwrite(mask_save_name, mask_png);
+                }
+                continue;
+            }
+
             for (int i = 0; i < batch_size; ++i) {
                 float* output_addr = reinterpret_cast<float*>(
                                     _outputs[0].data.data())
-                                   + i * (out_num / batch_size);
+                                   + i * (nums / batch_size);
                 output_mask(imgs_batch[i], output_addr,
-                            out_num / batch_size,
+                            nums / batch_size,
                             &org_height[i],
                             &org_width[i]);
             }
@@ -278,8 +305,14 @@ namespace PaddleSolution {
                 return -1;
             }
             auto im_tensor = _main_predictor->GetInputTensor("image");
-            im_tensor->Reshape({ batch_size, channels,
+            if (!_model_config._use_pr) {
+                im_tensor->Reshape({ batch_size, channels,
                                  eval_height, eval_width });
+            } else {
+                im_tensor->Reshape({ batch_size, eval_height,
+                                eval_width, channels});
+            }
+
             im_tensor->copy_from_cpu(input_buffer.data());
 
             auto t1 = std::chrono::high_resolution_clock::now();
@@ -292,7 +325,6 @@ namespace PaddleSolution {
             auto output_names = _main_predictor->GetOutputNames();
             auto output_t = _main_predictor->GetOutputTensor(
                                               output_names[0]);
-            std::vector<float> out_data;
             std::vector<int> output_shape = output_t->shape();
 
             int out_num = 1;
@@ -303,6 +335,30 @@ namespace PaddleSolution {
             }
             std::cout << ")" << std::endl;
 
+            if (_model_config._use_pr) {
+                    std::vector<int64_t> out_data;
+                    out_data.resize(out_num);
+                    output_t->copy_to_cpu(out_data.data());
+
+                    std::vector<uchar> mask_data;
+                    mask_data.resize(out_num);
+                    auto addr = reinterpret_cast<int64_t*>(out_data.data());
+                    for (int r = 0; r < out_num; ++r) {
+                        mask_data[r] = (int)(addr[r]);
+                    }
+                    for (int r = 0; r < batch_size; ++r) {
+                        cv::Mat mask_png = cv::Mat(eval_height, eval_width, CV_8UC1);
+                        mask_png.data = mask_data.data() + eval_height*eval_width*r;
+                        auto name = imgs_batch[r];
+                        auto pos = name.rfind(".");
+                        name[pos] = '_';
+                        std::string mask_save_name = name + "_mask.png";
+                        cv::imwrite(mask_save_name, mask_png);
+                    }
+                    continue;
+            }
+
+            std::vector<float> out_data;
             out_data.resize(out_num);
             output_t->copy_to_cpu(out_data.data());
             for (int i = 0; i < batch_size; ++i) {
diff --git a/deploy/cpp/preprocessor/preprocessor_seg.cpp b/deploy/cpp/preprocessor/preprocessor_seg.cpp
index 7c74042071143be53e11f1e7915531ea9354f356..11505a139046603b3c25bcd20af151eb871af4be 100644
--- a/deploy/cpp/preprocessor/preprocessor_seg.cpp
+++ b/deploy/cpp/preprocessor/preprocessor_seg.cpp
@@ -40,14 +40,18 @@ namespace PaddleSolution {
             LOG(ERROR) << "Only support rgb(gray) and rgba image.";
             return false;
         }
-
         cv::Size resize_size(_config->_resize[0], _config->_resize[1]);
         int rw = resize_size.width;
         int rh = resize_size.height;
         if (*ori_h != rh || *ori_w != rw) {
             cv::resize(im, im, resize_size, 0, 0, cv::INTER_LINEAR);
         }
-        utils::normalize(im, data, _config->_mean, _config->_std);
+
+        if (!_config->_use_pr) {
+            utils::normalize(im, data, _config->_mean, _config->_std);
+        } else {
+            utils::flatten_mat(im, data);
+        }
         return true;
     }
 
diff --git a/deploy/cpp/utils/seg_conf_parser.h b/deploy/cpp/utils/seg_conf_parser.h
index c217078c84421fa0e292110bf78369caa8c32a07..080a0530f440e12459644f771a7b4e07641323d2 100644
--- a/deploy/cpp/utils/seg_conf_parser.h
+++ b/deploy/cpp/utils/seg_conf_parser.h
@@ -25,6 +25,7 @@ class PaddleSegModelConfigPaser {
         :_class_num(0),
         _channels(0),
         _use_gpu(0),
+        _use_pr(0),
         _batch_size(1),
         _model_file_name("__model__"),
         _param_file_name("__params__") {
@@ -40,6 +41,7 @@ class PaddleSegModelConfigPaser {
         _class_num = 0;
         _channels = 0;
         _use_gpu = 0;
+        _use_pr = 0;
         _batch_size = 1;
         _model_file_name.clear();
         _model_path.clear();
@@ -172,6 +174,12 @@ class PaddleSegModelConfigPaser {
             std::cerr << "Please set CHANNELS: x"  << std::endl;
             return false;
         }
+        // 15. use_pr
+        if (config["DEPLOY"]["USE_PR"].IsDefined()) {
+            _use_pr = config["DEPLOY"]["USE_PR"].as<int>();
+        } else {
+            _use_pr = 0;
+        }
         return true;
     }
 
@@ -238,6 +246,8 @@ class PaddleSegModelConfigPaser {
     std::string _predictor_mode;
     // DEPLOY.BATCH_SIZE
     int _batch_size;
+    // USE_PR: OP Optimized model
+    int _use_pr;
 };
 
 }  // namespace PaddleSolution
diff --git a/deploy/cpp/utils/utils.h b/deploy/cpp/utils/utils.h
index 7e322daa03c02e704509f032d5709684a341060f..4578a6a8a612c955564cfb6ac279b0f92c9f3923 100644
--- a/deploy/cpp/utils/utils.h
+++ b/deploy/cpp/utils/utils.h
@@ -23,7 +23,8 @@
 #include <opencv2/highgui/highgui.hpp>
 
 #ifdef _WIN32
-#include <filesystem>
+#define GLOG_NO_ABBREVIATED_SEVERITIES
+#include <windows.h>
 #else
 #include <dirent.h>
 #include <sys/types.h>
@@ -67,15 +68,21 @@ namespace utils {
     // scan a directory and get all files with input extensions
     inline std::vector<std::string> get_directory_images(
                     const std::string& path, const std::string& exts) {
+        std::string pattern(path);
+        pattern.append("\\*");
         std::vector<std::string> imgs;
-        for (const auto& item :
-                    std::experimental::filesystem::directory_iterator(path)) {
-            auto suffix = item.path().extension().string();
-            if (exts.find(suffix) != std::string::npos && suffix.size() > 0) {
-                auto fullname = path_join(path,
-                                          item.path().filename().string());
-                imgs.push_back(item.path().string());
-            }
+        WIN32_FIND_DATA data;
+        HANDLE hFind;
+        if ((hFind = FindFirstFile(pattern.c_str(), &data)) != INVALID_HANDLE_VALUE) {
+            do {
+                auto fname = std::string(data.cFileName);
+                auto pos = fname.rfind(".");
+                auto ext = fname.substr(pos + 1);
+                if (ext.size() > 1 && exts.find(ext) != std::string::npos) {
+                    imgs.push_back(path + "\\" + data.cFileName);
+                }
+            } while (FindNextFile(hFind, &data) != 0);
+            FindClose(hFind);
         }
         return imgs;
     }
@@ -103,6 +110,25 @@ namespace utils {
         }
     }
 
+    // flatten a cv::mat
+    inline void flatten_mat(cv::Mat& im, float* data) {
+        int rh = im.rows;
+        int rw = im.cols;
+        int rc = im.channels();
+        #pragma omp parallel for
+        for (int h = 0; h < rh; ++h) {
+            const uchar* ptr = im.ptr<uchar>(h);
+            int im_index = 0;
+            int top_index = h * rw * rc;
+            for (int w = 0; w < rw; ++w) {
+                for (int c = 0; c < rc; ++c) {
+                    float pixel = static_cast<float>(ptr[im_index++]);
+                    data[top_index++] = pixel;
+                }
+            }
+        }
+    }
+
     // argmax
     inline void argmax(float* out, std::vector<int>& shape,
                        std::vector<uchar>& mask, std::vector<uchar>& scoremap) {
diff --git a/pdseg/export_model.py b/pdseg/export_model.py
index 410cfc33432255fda8ccc5fada1ec0433f11f243..93e4b4949fa3f18c934844edbad4693813e67f15 100644
--- a/pdseg/export_model.py
+++ b/pdseg/export_model.py
@@ -52,6 +52,7 @@ def parse_args():
 def export_inference_config():
     deploy_cfg = '''DEPLOY:
         USE_GPU : 1
+        USE_PR : 1
         MODEL_PATH : "%s"
         MODEL_FILENAME : "%s"
         PARAMS_FILENAME : "%s"