嵌入式开发板上部署目标检测模型mobilenetv2_ssdlite

本文链接：https://blog.csdn.net/zhongqianli/article/details/110450557

文章目录

项目地址

mobilenetv2_ssdlite_ncnn

目标

在RK3308嵌入式开发板（Arm Linux）上使用ncnn部署mobilenetv2_ssdlite模型
模型量化

细节

模型量化

https://github.com/Tencent/ncnn/tree/master/tools/quantize

优化图

https://github.com/Tencent/ncnn/wiki/use-ncnnoptimize-to-optimize-model

./ncnnoptimize mbv2_ssdlite/mobilenetv2_ssdlite_voc.param mbv2_ssdlite/mobilenetv2_ssdlite_voc.bin mbv2_ssdlite_nobn/mobilenetv2_ssdlite_voc.param mbv2_ssdlite_nobn/mobilenetv2_ssdlite_voc.bin 0

算子合并

batchnorm - scale
convolution - batchnorm
convolutiondepthwise - batchnorm
deconvolution - batchnorm
deconvolutiondepthwise - batchnorm
innerproduct - batchnorm
convolution - relu
convolutiondepthwise - relu
deconvolution - relu
deconvolutiondepthwise - relu
innerproduct - relu

消除空算子

innerproduct - dropout
flatten after global pooling

偏好更好的算子

replace convolution with innerproduct after global pooling

创建校准表文件

我们建议使用验证数据集进行校准，该数据集应超过5000张图像。

ncnn2table --param=mbv2_ssdlite_nobn/mobilenetv2_ssdlite_voc.param --bin=mbv2_ssdlite_nobn/mobilenetv2_ssdlite_voc.bin --images=../images/ --output=mbv2_ssdlite_nobn.table --mean=127.5,127.5,127.5 --norm=0.007843,0.007843,0.007843 --size=300,300 --thread=10

量化

ncnn2int8 mbv2_ssdlite_nobn/mobilenetv2_ssdlite_voc.param mbv2_ssdlite_nobn/mobilenetv2_ssdlite_voc.bin mbv2_ssdlite_int8/mobilenetv2_ssdlite_voc.param mbv2_ssdlite_int8/mobilenetv2_ssdlite_voc.bin mbv2_ssdlite_nobn.table

修改输入分辨率

修改模型文件mobilenetv2_ssdlite_voc.param的分辨率为192x192，原分辨率为300x300
```
Input                    data                     0 1 data 0=192 1=192 2=3
```

修改mobilenetv2ssdlite.cpp中的输入图像的分辨率为192x192，原分辨率为300x300

    const int target_size = 192;

    int img_w = bgr.cols;
    int img_h = bgr.rows;

    ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, target_size, target_size);