一种改进型级联神经网络检测算法及加速处理
An Improved Cascaded Neural Network Detection Algorithm And Accelerated Processing
张子振 1南钢洋 1孟凡超 1白雪1
作者信息
- 1. 齐鲁工业大学(山东省科学院)激光研究所,山东济南,250104
- 折叠
摘要
为提高MTCNN网络检测准确度,且针对检测密集样本容易漏检的问题,通过改进网络隐藏层结构提高网络学习能力,通过Soft-NMS惩罚置信度方式筛选检测框,提高了网络检测准确度,针对密集样本仍保持高精度;且为提高改进后网络推理速度和克服网络依赖PC端资源问题,基于HLS实现了网络加速推理.实验结果表明,改进后各子网络检测准确度由93.73%、95.30%、95.89%提高至94.78%、96.30%、97.55%,密集样本测试集测试准确度为 97.21%;使用硬件加速对比2.9GhzCPU推理速度快3.3 倍,硬件资源最大占用91%,较好利用硬件资源实现了加速处理.
Abstract
In order to improve the detection accuracy of the MTCNN network detection and address the problem of missed detections in detecting dense samples,this paper improved the network's learning ability by improving the hid-den layer structure,and filters detection boxes through the Soft-NMS penalty confidence method to improve the accu-racy of network detection.High accuracy is still maintained for dense samples.And in order to improve the network reasoning speed and overcome the problem of network dependence on PC resources,this paper implemented network accelerated reasoning based on HLS.The experimental results show that the detection accuracy of each subis increased from 93.73%,95.3%and 95.89%to 94.78%,96.3%and 97.55%,and the test accuracy of the dense sample test set is 97.21%;Compared with 2.9Ghz CPU,the reasoning speed is 3.3 times faster by using hardware acceleration,and the maximum of hardware resources are 91%.The hardware resources are better used to realize ac-celerated processing.
关键词
神经网络/人脸检测/置信度/推理加速Key words
Neural network/Face detection/Confidence/Inference acceleration引用本文复制引用
基金项目
济南市高等学校20条自主培养创新团队项目(2020GXRC004)
齐鲁工业大学(山东省科学院)计算机科学与技术学科基础研究加强项目(2021JC02008)
出版年
2024
NETL
NSTL
万方数据