基于YOLO框架的轻量化高精度目标检测算法

Lightweight and high-precision object detection algorithm based on YOLO framework

面对图像的多尺度目标检测算法存在检测精度和系统开销相互制约的问题,提出了一种基于YOLO框架的轻量化高精度目标检测算法。在YOLO框架下,以MobileNetv3网络为基础,改进下采样和通道注意力机制,以精准提取目标特征,降低了不必要开销。设计了特征金字塔和单阶段无头融合的结构,构建了不同感受野以获取不同尺度信息,增强了算法对多尺度目标的适应性。同时采用SIOU作为回归损失和Soft-NMS进行冗余框处理,提高了算法的精度。在MS COCO数据集和UA-DETRAC交通监控数据集上的实验结果表明,本文提出的改进算法与原YOLOXs相比,在不降低精度的情况下,模型参数量和计算量分别降低了64.98%、57.14%;在UA-DETRAC交通数据集上,mAP@0.5达到70.5%,提升了3.52%,FPS提升了14.4%。本文改进算法大幅降低了系统开销,提升了精度,有效保障了检测的双重性能。

Image-oriented multi-scale object detection algorithms often have the problem of mutual restriction between detection accuracy and system cost.Therefore,a lightweight and high-precision object detection algorithm based on YOLO framework is proposed.Under the YOLO framework,the mechanism of down-sampling and channel attention based on MobileNetv3 network is improved to accurately extract target features and reduce unnecessary overhead.The feature pyramid and single-stage headless fusion structure are designed,and different receptive fields are constructed to obtain different scale information,so as to enhance the adaptability of the algorithm for multi-scale targets.At the same time,SIOU is used as regression loss and Soft-NMS is used for redundant frame processing to improve the accuracy of the algorithm.Experiments are conducted on the MS COCO and UA-DETRAC.Compared with the original YOLOXs,the results show that the proposed improved algorithm reduces the number of model parameters and the computational cost reduced by 64.98%and 57.14%without reducing the accuracy.On the UA-DETRAC,mAP@0.5 reaches 70.5%which is improved by 3.52%,and FPS increases by 14.4%.The experimental results show that our algorithm greatly reduces the system overhead,improves the accuracy,and effectively guarantees the dual performance of detection.