一种改进的MTCNN人脸检测算法

人脸检测主要运用于机场、火车站等人口密集场所。目前常用的人脸检测算法有MTCNN、YOLOV3、faster-RCNN、SSD等，但已有算法难以兼顾检测速度和检测准确性。 通过改进多任务级联卷积神经网络（MTCNN）人脸检测算法，将MTCNN、YOLOV3、faster-RCNN等3个模型进行整合，减少内存与显存之间的数据搬运；然后动态修改Minsize值，减少图像金字塔中图片生成数量，并根据图像相似度对输入图片进行不同处理以提高效率。改进后的MTCNN算法比传统MTCNN算法识别速度提高将近40%，且正确率达到97%，可更好满足现代社会对于人脸检测的应用需求。

A Face Detection Algorithm Base on Improved MTCNN

Face detection is mainly used in highly-crowded areas， such as airports， railway stations and so on. At present， the popular face detection algorithms are MTCNN， YOLOV3， fast-RCNN， SSD ，etc. However none of the existed algorithms excel both in detection speed and accuracy. It is difficult to find a fast and accurate detection results algorithm on the market. By improving the MTCNN face detection algorithm， the three models of MTCNN， YOLOV3， and faster-RCNN are integrated to reduce the data handling between memory and video memory； the number of image generation in the image pyramid is reduced by dynamically modify the Minsize value； and based on image similarity the input picture is processed differently to improve the efficiency. The improved MTCNN algorithm is nearly 40% faster than the traditional MTCNN algorithm， and the correct rate is 97%， which can better meet the needs of modern society for face detection.