[1]陈钧,周井泉*,张志鹏.面向小目标的自校正YOLOv5检测增强算法[J].计算机技术与发展,2024,34(10):140-147.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0220]
　CHEN Jun,ZHOU Jing-quan*,ZHANG Zhi-peng.Self-calibrated YOLOv5 Detection Enhancement Algorithm for Small Targets[J].,2024,34(10):140-147.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0220]
点击复制

面向小目标的自校正YOLOv5检测增强算法()

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

卷:
34
期数:
2024年10期
页码:
140-147
栏目:
人工智能
出版日期:
2024-10-10

文章信息/Info

Title:
Self-calibrated YOLOv5 Detection Enhancement Algorithm for Small Targets
文章编号:
1673-629X(2024)10-0140-08
作者:
陈钧; 周井泉*; 张志鹏
南京邮电大学 电子与光学工程学院、柔性电子学院,江苏 南京 210023
Author(s):
CHEN Jun; ZHOU Jing-quan*; ZHANG Zhi-peng
School of Electronic and Optical Engineering,School of Flexible Electronics,Nanjing University of Posts and Telecommunications,Nanjing 210023,China
关键词:
智能驾驶; 自校正卷积; YOLOv5s; 轻量化; 小目标检测
Keywords:
intelligent driving; self-calibrated convolutions; YOLOv5s; lightweight; small targets detection
分类号:
TP391;TP183
DOI:
10.20165/j.cnki.ISSN1673-629X.2024.0220
摘要:
为了解决汽车智能驾驶精度不够高、小目标漏检和误检的问题,提出了一种基于自校正卷积的 YOLOv5s 的道路目标检测算法。 该算法主要设计了一种自校正卷积网络,通过深层特征提取以及特征融合来提高检测精度以及对小目标的检测能力。 对自校正卷积网络进行轻量化处理,减少模型的大小以及训练过程中的参数量。 增加小目标校正检测层,输出检测小目标的特征图。 设计了 SAIoU 损失函数来代替目标框回归中的 CIoU 损失函数,加速目标框的回归。 在公开的自动驾驶 KITTI 数据集和 BDD100K 数据集上对该算法进行了测试,检测的平均精度均值( mAP) 分别可达到 95. 1% 和53.1% ,相比于 YOLOv5s 算法分别提高了 2. 1 百分点和 4. 2 百分点。 与其他算法进行对比实验,在检测精度和小目标检测能力上具有一定的优势,并且轻量化的自校正卷积网络在模型大小上相比于未进行轻量化处理的自校正卷积网络压缩了 35% ,提高了实时性。 结果表明该算法能够满足实时性,能够提升检测精度以及小目标检测能力。
Abstract:
In order to solve the problems of low precision of intelligent driving and missing and false detection of small targets,a road target detection algorithm based on self-calibrated convolutional YOLOv5s is proposed. In this algorithm,a self-calibrated convolutions network is designed to improve the detection accuracy and detection ability of small targets through deep feature extraction and feature fusion. Lightweight processing is applied to the self-calibrated convolutions network to reduce model size and the number of parameters during training. A small target calibration detection layer is added to output feature maps for detecting small objects. Additionally,the SAIoU loss function is designed to replace the CIoU loss function in target box regression,accelerating the regression of target boxes. The proposed algorithm was tested on KITTI and BDD100K publicly available autopilot datasets,and the mean average precision of detection can reach 95. 1% and 53. 1% ,respectively. Compared to the YOLOv5s algorithm,it improves by 2. 1 and 4. 2 percentage points on these datasets,respectively. Comparative experiments with other algorithms demonstrate certain advantages in detection accuracy and small targets detection capability. Moreover,the lightweight self-calibrated convolutions network compresses the model size by 35% compared to its non-lightweight,enhancing real-time performance. These results indicate that the proposed algorithm meets real-time requirements and can improve detection accuracy and small targets detection capability.

相似文献/References:

[1]岳江,刘庆晨,韩晓鑫,等.基于自校正卷积与注意力的脑白质病变分割[J].计算机技术与发展,2024,34(12):25.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0264]
　YUE Jiang,LIU Qing-chen,HAN Xiao-xin,et al.Segmentation of Brain White Matter Lesions Using Fusion of Self-calibrated Convolution and Attention[J].,2024,34(10):25.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0264]

更新日期/Last Update: 2024-10-10