【深度学习】YOLOv4: Optimal Speed and Accuracy of Object Detection 目标检测

YOLOv4论文:https://arxiv.org/pdf/2004.10934.pdf

yolov4 技术

(1)Weighted-Residual-Connections (WRC)
https://arxiv.org/pdf/1605.08831.pdf

(2)Cross-Stage-Partial-connections (CSP) backbone
https://blog.csdn.net/x1131230123/article/details/123735975

(3)Cross mini-Batch Normalization (CmBN):
https://arxiv.org/pdf/2002.05712.pdf
BN:对当前mini-batch进行归一化
CBN: 对当前以及当前往前数3个mini-batch的结果进行归一化
CmBN: CmBN 在整个批次中使用Cross min-batch Normalization 收集统计数据,而非在单独的mini-batch中收集统计数据。

(4) Self-adversarial-training (SAT)
https://arxiv.org/pdf/1707.02439.pdf

(5)Mish-activation
https://arxiv.org/pdf/1908.08681.pdf
参考https://github.com/digantamisra98/Mish

(6)Mosaic data augmentation

(7)DropBlock regularization 在CNN特征图dropout
https://arxiv.org/abs/1810.12890

(8)CIoU loss
https://blog.csdn.net/x1131230123/article/details/123751062

(9)PANnet neck
https://blog.csdn.net/x1131230123/article/details/123882314

(10)SPPnet neck
https://arxiv.org/pdf/1406.4729.pdf
【深度学习】YOLOv4: Optimal Speed and Accuracy of Object Detection 目标检测_第1张图片

【深度学习】YOLOv4: Optimal Speed and Accuracy of Object Detection 目标检测_第2张图片

yolov5 损失函数

【深度学习】YOLOv4: Optimal Speed and Accuracy of Object Detection 目标检测_第3张图片

你可能感兴趣的:(深度学习机器学习,深度学习)