【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第1张图片

Neural Computing and Applications 2023


文章目录

  • 1 Background and Motivation
  • 2 Related Work
  • 3 Advantages / Contributions
  • 4 Method
    • 4.1 The improved YOLOv5s network framework
    • 4.2 AF-FPN structure
    • 4.3 Data augmentation
  • 5 Experiments
    • 5.1 Datasets and Metrics
    • 5.2 Experimental analysis
  • 6 Conclusion(own)


1 Background and Motivation

基于深度学习方法的交通标志检测,面临 multi-scale targets and the real-time problem,作者基于 yolov5 中尺寸最小的 yolov5s 改进,使其比 SOTA more universal and superior.

2 Related Work

  • CNN-based traffic sign detection
  • Data augmentation(color transformation / geometric transformation)

3 Advantages / Contributions

改进 yolov5,使其更好更快的识别出交通标志

  • 基于特征金字塔,提出 AF-FPN——新引入 adaptive attention module(AAM) 和 feature enhancement module(FEM) 两个模块
  • 数据增强方法改成了 automatic learning data augmentation method

4 Method

high-level features will lose a lot of information, resulting in a decrease in
the detection accuracy of large-scale targets

4.1 The improved YOLOv5s network framework

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第2张图片
整体结构,创新处用绿色√标注

损失用的 CIoU,直接引用就好了,还来个 GIoU 的公式说不好,再列出 CIoU,这。。。

后处理用的是 softer NMS

4.2 AF-FPN structure

yolov5 特征金字塔中引入下面两个模组

  • adaptive attention module (AAM)

  • and the feature enhancement module (FEM)

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第3张图片

(1)AAM

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第4张图片
作用:reduces the loss of context information in the high-level feature map due to the reduced feature channels

pooling 成不同分辨率,再上采样回来,紧接着一个空间注意力,最后和输入加起来为 M6,which contains multi-scale context information.

有没有想过,输入的 C5 本身就丢失的比较多,空间信息回不来了呢?还不如 adaptive pooling layer 直接接金字塔(C3 / C4 / C5)而不仅仅是 C5

(2)FEM

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第5张图片
作用:enhances the representation of feature pyramids and accelerates the inference speed while achieving state-of-the-art performance

空洞卷积,最后 pooling 合起来,公式如下

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第6张图片
B = 3,公式是加起来了

如果 pooling 核一样,配合 dilation rate 还行

如果 pooling 核不一样,同 yolov5 的 SPP,那和 dilation 搭配就有点怪怪的感觉,没有看代码

4.3 Data augmentation

mosaic data augmentation 的缺点

it will cause the original small targets in the dataset to become smaller, resulting in the deterioration of the generalization performance of the model.

作者弃用了 mosaic data augmentation

Auto data augmentation 那一套,没有什么可以说的

【AutoAugment】《AutoAugment:Learning Augmentation Policies from Data》

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第7张图片
强化学习搜索空间 ( 19 × D × P ) 2 × 5 (19 \times D \times P)^{2 \times 5} (19×D×P)2×5

  • 19 总增广方法

  • 增广强度 D 离散化为 11 个等级

  • 增广概率 P 离散化为 10 个等级

  • 5 个增广组合,每个组合抽 2 种增广方法

5 Experiments

5.1 Datasets and Metrics

TT100K

182 types of traffic signs
【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第8张图片
目标大小分布

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第9张图片
中小目标居多

评价指标

  • mAP
  • log-average miss rate(LAMR,平均对数漏检率): Calculated by averaging miss rates at 9 evenly spaced FPPI points between 10e-2 and 10e0, in log-space——《Pedestrian Detection: An Evaluation of the State of the Art》(PAMI-2011)

其中 FPPI:false positive per image,每张图片的平均误检率
F P P I = F P N FPPI = \frac{FP}{N} FPPI=NFP

M R = F N G T = 1 − R e c a l l MR = \frac{FN}{GT} = 1 - Recall MR=GTFN=1Recall

MR-FPPI 与目标检测所用的 Precious-Recall 类似

采用 FPPI 为横坐标,MR 为纵坐标的曲线中,均匀选取 范围内的 9 个FPPI,得到他们对应的9个值,并对这几个纵坐标值进行平均,最后通过指数运算上述平均值恢复为MR的百分比形式,就获得了用来量化MR-FPPI 曲线的指标,该指标越小代表检测器性能越高(来自深度学习评估指标之目标检测——(yolov5 可视化训练结果以及result.txt解析) )

5.2 Experimental analysis

608x608 输入
train+val 9146 images
test 1121 image

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第10张图片
要速度没有速度,要精度没有精度

小目标( A P s AP_s APs)确实有提升

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第11张图片

这样看,YOLOv5-Face 效果最好,哈哈哈

效果展示

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第12张图片

消融实验

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第13张图片

+Aug Model 变大了?训练出来的文件吧,权重变化了,统计这个有啥用啊

【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第14张图片

6 Conclusion(own)

  • 好久没有写论文博客了,还是要坚持啃啃论文啦

  • 堆参数而以,引进的结构有 【BiSeNet】《BiSeNet:Bilateral Segmentation Network for Real-time Semantic Segmentation》 的影子,AAM 改进改进可能更合理(合我的理,哈哈),不过未必能提点

  • intelligent transportation systems (ITS)

  • The effect of data augmentation strategies is related to the characteristics of the dataset itself

  • 对抗学习来提升小目标,《Perceptual Generative Adversarial Networks for Small Object Detection》(CVPR-2017)

    【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第15张图片
    判别器判断是真实的大目标还是伪装成大目标的小目标
    【Yolov5 Traffic Sign】《Improved YOLOv5 network for real-time multi-scale traffic sign detection》_第16张图片

你可能感兴趣的:(CNN,Traffic,Sign,yolov5,AF-FPN,AAM,FEM)