[1]朱 郝,杨世恩,陈春梅.基于结构重参数化技术的轻量化目标检测算法[J].计算机技术与发展,2023,33(10):169-175.[doi:10. 3969 / j. issn. 1673-629X. 2023. 10. 026]
　ZHU Hao,YANG Shi-en,CHEN Chun-mei.A Lightweight Object Detection Algorithm Based on Structural Re-parameterization Technique[J].,2023,33(10):169-175.[doi:10. 3969 / j. issn. 1673-629X. 2023. 10. 026]

点击复制

基于结构重参数化技术的轻量化目标检测算法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年10期

页码:

169-175

栏目:

人工智能

出版日期:

2023-10-10

文章信息/Info

Title:

A Lightweight Object Detection Algorithm Based on Structural Re-parameterization Technique

文章编号:

1673-629X(2023)10-0169-07

作者:

朱 郝¹ ; 杨世恩¹ ; 陈春梅²

1. 西南科技大学 计算机科学与技术学院,四川 绵阳 621010;

2. 西南科技大学 信息工程学院,四川 绵阳 621010

Author(s):

ZHU Hao1 ; YANG Shi-en1 ; CHEN Chun-mei2

1. School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,China;

2. School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China

关键词:

目标检测; 结构重参数化; MobileOne; YOLO; 轻量化网络

Keywords:

object detection; structural re-parameterization; MobileOne; YOLO; lightweight network

分类号:

TP399

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 10. 026

摘要:

现有目标检测算法消耗大量算力资源、参数量大、占用内存空间多,不利于在小型设备上推广使用。 因此,基于结构重参数化技术并结合 YOLO 系列算法的研究成果,提出了一种轻量化目标检测模型 Rep-YOLO。 使用结构重参数化技术实现模型在训练时的多分支结构和推理时的线性结构之间的转换,从而减少模型推理时对算力资源的消耗。 另外,为了降低模型的参数量,利用深度可分离卷积、网络裁剪等方法,重新设计了多尺度特征融合网络和检测头。 实验结果表明:在 PASCAL VOC 上,Rep-YOLO-s1 精度可达 82. 7% ,Rep-YOLO-s1 与 YOLOv6s 相比,在参数量减少 54. 8% 的情况下,精度提高了 2. 4 百分点,推理速度提升了 6% 。 在 NVIDIA RTX 3060 GPU 上,Rep -YOLO-s0 比 YOLOv6s 的推理速度快10% ,Rep-YOLO-nano 比 YOLOv7-tiny 快 4% ,精度提高了 0. 5 百分点。 Rep-YOLO 与规模类似的模型相比,体积更小,精度更高,更加利于资源有限的部署应用。

Abstract:

The existing target detection algorithm consumes a lot of computing resources,has a large number of parameters and occupies alot of memory space, which is not conducive to the promotion of use on small devices. Therefore, based on the structural re -parameterization technique and combined with the research results of YOLO series algorithms,a lightweight target detection model Rep-YOLO is proposed. Such decoupling of the multi-branch structure at training time and the typical linear structure at inference time isachieved by a structural re - parameterization technique, thus reducing the consumption of computing resources at inference time. Inaddition,the multi-scale feature fusion network and detection head are redesigned using depthwise separable convolution and networkcropping in order to reduce?

the number of parameters in the model. The experimental results show that on PASCAL VOC,the accuracy of　Rep-YOLO-s1 can reach 82. 7% . Compared with YOLOv6s,the accuracy of Rep-YOLO-s1 is increased by 2. 4 percentage points andthe reasoning speed is increased by 6% when the number of parameters is reduced by 54. 8% . On the NVIDIA RTX 3060 GPU,the Rep-YOLO-s0 inference is 10% faster than the YOLOv6s,and the Rep - YOLO - nano is 4% faster than the YOLOv7 - tiny, with a 0. 5percentage point improvement in accuracy. Compared with similar - scale models, Rep - YOLO is smaller, more accurate, and moreconducive to resource-limited deployment applications.

常用功能