MTCNN-PureC

MTCNN face detection with C and Arm-CMSIS-DSP

Introduction

The project is a MTCNN face detection based on C. To implement MTCNN in MCU, the project doesn't rely on any library and you can use ARM-CMSIS-DSP to accelerate the matrix multiplication. The input data is a RGB CHW txt, and output is the keypoints of face and bounding box of face.

本项目为MTCNN的C实现，无依赖库，主要为arm平台的mcu编写，在部署时可通过ARM-CMSIS-DSP库来加速矩阵乘法运算，请参考官方文档。

本项目在保持API不变下简化了arm_math.h和arm_mat_mult_f32.c，以便在PC平台运行。

输入为RGB CHW格式的文本文件，输出为人脸包围盒和人脸关键点。

Usage

git clone https://github.com/notplus/MTCNN-PureC.git
cd MTCNN-PureC
mkdir build
cd build
cmake ..
make

Demo

Input

输入为RGB CHW 格式图像生成的文本文件，可使用如下代码生成，另外使用时需要在include/test.h中修改图像尺寸。

import numpy as np
import cv2
img = cv2.imread("1.jpg")
img_ = img[:,:,::-1].transpose((2,0,1))
np.savetxt('input1.txt', img_.reshape(-1), "%d")

Output

• boundingBox3 is 160 72 450 362 //(x1, y1, x2, y2)
• keyPoint0 is 259.010193 188.661850
• keyPoint1 is 350.958679 183.481461
• keyPoint2 is 309.558563 237.440964
• keyPoint3 is 267.242126 291.008728
• keyPoint4 is 354.936646 283.069946

Print with opencv

Implement Details

The project mainly refers to MTCNN-light and remove OpenCV and OpenBLAS dependencies.
Convolution is implemented by im2col and matrix multiplication.

本项目主要参考了MTCNN-light，并且移除了OpenCV和OpenBLAS依赖。