深圳幻海软件技术有限公司 欢迎您！
更多

YOLOv8/YOLOv7/YOLOv5系列算法改进【NO.6】增加小目标检测层，提高对小目标的检测效果

2023-03-25

none conv cat
前言：作为当前先进的深度学习目标检测算法YOLO，已经集合了大量的trick，但是在处理一些复杂检测问题的时候，还是容易出现错漏检的问题。此后的系列文章，将重点对YOLOv8、YOLOv7以及YOLOv5的如何改进进行详细的介绍，目的是为了给那些搞科研的同学需要创新点或者搞工程项目的朋友需要达到更好

前 言：作为当前先进的深度学习目标检测算法YOLO，已经集合了大量的trick，但是在处理一些复杂检测问题的时候，还是容易出现错漏检的问题。此后的系列文章，将重点对YOLOv8、YOLOv7以及YOLOv5的如何改进进行详细的介绍，目的是为了给那些搞科研的同学需要创新点或者搞工程项目的朋友需要达到更好的效果提供自己的微薄帮助和参考。

一、解决问题

        YOLO小目标检测效果不好的一个原因是因为小目标样本的尺寸较小，而yolov8的下采样倍数比较大，较深的特征图很难学习到小目标的特征信息，因此提出增加小目标检测层对较浅特征图与深特征图拼接后进行检测。加入小目标检测层，可以让网络更加关注小目标的检测，提高检测效果。这个方式的实现十分简单有效，只需要修改yolov8的模型文件yaml就可以增加小目标检测层，但是在增加检测层后，带来的问题就是计算量增加，导致推理检测速度降低。不过对于小目标，确实有很好的改善，修改yaml文件，需要修改特征融合网络。

二、YOLOv8改进方法

近期有朋友问到YOLOv8的改进方法，特此分享，增加小目标检测层的yaml文件前后对比。

yaml文件

改进前：

  1.  
  2. # Ultralytics YOLO 🚀, GPL-3.0 license
  3.  
  4. # Parameters
  5. nc: 80  # number of classes
  6. depth_multiple: 0.33  # scales module repeats
  7. width_multiple: 0.50  # scales convolution channels
  8.  
  9. # YOLOv8.0s backbone
  10. backbone:
  11.   # [from, repeats, module, args]
  12.   - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
  13.   - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
  14.   - [-1, 3, C2f, [128, True]]
  15.   - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
  16.   - [-1, 6, C2f, [256, True]]
  17.   - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
  18.   - [-1, 6, C2f, [512, True]]
  19.   - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
  20.   - [-1, 3, C2f, [1024, True]]
  21.   - [-1, 1, SPPF, [1024, 5]]  # 9
  22.  
  23. # YOLOv8.0s head
  24. head:
  25.   - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  26.   - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
  27.   - [-1, 3, C2f, [512]]  # 13
  28.  
  29.   - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  30.   - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
  31.   - [-1, 3, C2f, [256]]  # 17 (P3/8-small)
  32.  
  33.   - [-1, 1, Conv, [256, 3, 2]]
  34.   - [[-1, 12], 1, Concat, [1]]  # cat head P4
  35.   - [-1, 3, C2f, [512]]  # 20 (P4/16-medium)
  36.  
  37.   - [-1, 1, Conv, [512, 3, 2]]
  38.   - [[-1, 9], 1, Concat, [1]]  # cat head P5
  39.   - [-1, 3, C2f, [1024]]  # 23 (P5/32-large)
  40.  
  41.   - [[15, 18, 21], 1, Detect, [nc]]  # Detect(P3, P4, P5)

改进后：

  1.  
  2. # Ultralytics YOLO 🚀, GPL-3.0 license
  3.  
  4. # Parameters
  5. nc: 80  # number of classes
  6. depth_multiple: 0.33  # scales module repeats
  7. width_multiple: 0.50  # scales convolution channels
  8.  
  9. # YOLOv8.0s backbone
  10. backbone:
  11.   # [from, repeats, module, args]
  12.   - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
  13.   - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
  14.   - [-1, 3, C2f, [128, True]]
  15.   - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
  16.   - [-1, 6, C2f, [256, True]]
  17.   - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
  18.   - [-1, 6, C2f, [512, True]]
  19.   - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
  20.   - [-1, 3, C2f, [1024, True]]
  21.   - [-1, 1, SPPF, [1024, 5]]  # 9
  22.  
  23. # YOLOv8.0s head
  24. head:
  25.   - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  26.   - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
  27.   - [-1, 3, C2f, [512]]  # 13
  28.  
  29.   - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  30.   - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
  31.   - [-1, 3, C2f, [256]]  # 17 (P3/8-small)
  32.  
  33.   - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  34.   - [[-1, 2], 1, Concat, [1]]  # cat backbone P3
  35.   - [-1, 3, C2f, [128]]  # 20 (P4/16-medium)
  36.  
  37.   - [-1, 1, Conv, [256, 3, 2]]
  38.   - [[-1, 15], 1, Concat, [1]]  # cat head P4
  39.   - [-1, 3, C2f, [256]]  # 20 (P4/16-medium)
  40.  
  41.   - [-1, 1, Conv, [512, 3, 2]]
  42.   - [[-1, 12], 1, Concat, [1]]  # cat head P5
  43.   - [-1, 3, C2f, [512]]  # 23 (P5/32-large)
  44.  
  45.   - [-1, 1, Conv, [512, 3, 2]]
  46.   - [[-1, 9], 1, Concat, [1]]  # cat head P5
  47.   - [-1, 3, C2f, [1024]]  # 23 (P5/32-large)
  48.  
  49.   - [[18, 21, 24,27], 1, Detect, [nc]]  # Detect(P3, P4, P5)

结构图

改进前：

 改进后：

 最后，将train.py中改为本文的yaml文件即可，开始训练。

三、YOLOv7改进方法

      对YOLOv7项目路径下cfg\deploy\yolov7-tiny-silu.yaml进行修改为[[74,75,76,77], 1, Detect, [nc, anchors]],   # Detect(P3, P4, P5)，同时锚框增加小目标检测头的锚框[5,6, 8,14, 15,11]。具体改进后的如下所示。

  1. # parameters
  2. nc: 80  # number of classes
  3. depth_multiple: 1.0  # model depth multiple
  4. width_multiple: 1.0  # layer channel multiple
  5.  
  6. # anchors
  7. anchors:
  8. #  - [10,13, 16,30, 33,23]  # P3/8
  9. #  - [30,61, 62,45, 59,119]  # P4/16
  10. #  - [116,90, 156,198, 373,326]  # P5/32
  11.   - [5,6, 8,14, 15,11]
  12.   - [10,13, 16,30, 33,23]  # P3/8
  13.   - [30,61, 62,45, 59,119]  # P4/16
  14.   - [116,90, 156,198, 373,326]  # P5/32
  15.  
  16. # YOLOv7-tiny backbone
  17. backbone:
  18.   # [from, number, module, args]
  19.   [[-1, 1, Conv, [32, 3, 2]],  # 0-P1/2  
  20.   
  21.    [-1, 1, Conv, [64, 3, 2]],  # 1-P2/4    
  22.    
  23.    [-1, 1, Conv, [32, 1, 1]],
  24.    [-2, 1, Conv, [32, 1, 1]],
  25.    [-1, 1, Conv, [32, 3, 1]],
  26.    [-1, 1, Conv, [32, 3, 1]],
  27.    [[-1, -2, -3, -4], 1, Concat, [1]],
  28.    [-1, 1, Conv, [64, 1, 1]],  # 7
  29.    
  30.    [-1, 1, MP, []],  # 8-P3/8
  31.    [-1, 1, Conv, [64, 1, 1]],
  32.    [-2, 1, Conv, [64, 1, 1]],
  33.    [-1, 1, Conv, [64, 3, 1]],
  34.    [-1, 1, Conv, [64, 3, 1]],
  35.    [[-1, -2, -3, -4], 1, Concat, [1]],
  36.    [-1, 1, Conv, [128, 1, 1]],  # 14
  37.    
  38.    [-1, 1, MP, []],  # 15-P4/16
  39.    [-1, 1, Conv, [128, 1, 1]],
  40.    [-2, 1, Conv, [128, 1, 1]],
  41.    [-1, 1, Conv, [128, 3, 1]],
  42.    [-1, 1, Conv, [128, 3, 1]],
  43.    [[-1, -2, -3, -4], 1, Concat, [1]],
  44.    [-1, 1, Conv, [256, 1, 1]],  # 21
  45.    
  46.    [-1, 1, MP, []],  # 22-P5/32
  47.    [-1, 1, Conv, [256, 1, 1]],
  48.    [-2, 1, Conv, [256, 1, 1]],
  49.    [-1, 1, Conv, [256, 3, 1]],
  50.    [-1, 1, Conv, [256, 3, 1]],
  51.    [[-1, -2, -3, -4], 1, Concat, [1]],
  52.    [-1, 1, Conv, [512, 1, 1]],  # 28
  53.   ]
  54.  
  55. # YOLOv7-tiny head
  56. head:
  57.   [[-1, 1, Conv, [256, 1, 1]],
  58.    [-2, 1, Conv, [256, 1, 1]],
  59.    [-1, 1, SP, [5]],
  60.    [-2, 1, SP, [9]],
  61.    [-3, 1, SP, [13]],
  62.    [[-1, -2, -3, -4], 1, Concat, [1]],
  63.    [-1, 1, Conv, [256, 1, 1]],
  64.    [[-1, -7], 1, Concat, [1]],
  65.    [-1, 1, Conv, [256, 1, 1]],  # 37
  66.   
  67.    [-1, 1, Conv, [128, 1, 1]],
  68.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  69.    [21, 1, Conv, [128, 1, 1]], # route backbone P4
  70.    [[-1, -2], 1, Concat, [1]],
  71.    
  72.    [-1, 1, Conv, [64, 1, 1]],
  73.    [-2, 1, Conv, [64, 1, 1]],
  74.    [-1, 1, Conv, [64, 3, 1]],
  75.    [-1, 1, Conv, [64, 3, 1]],
  76.    [[-1, -2, -3, -4], 1, Concat, [1]],
  77.    [-1, 1, Conv, [128, 1, 1]],  # 47
  78.   
  79.    [-1, 1, Conv, [64, 1, 1]],
  80.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  81.    [14, 1, Conv, [64, 1, 1]], # route backbone P3
  82.    [[-1, -2], 1, Concat, [1]],
  83.    
  84.    [-1, 1, Conv, [32, 1, 1]],
  85.    [-2, 1, Conv, [32, 1, 1]],
  86.    [-1, 1, Conv, [32, 3, 1]],
  87.    [-1, 1, Conv, [32, 3, 1]],
  88.    [[-1, -2, -3, -4], 1, Concat, [1]],
  89.    [-1, 1, Conv, [64, 1, 1]],  # 57
  90.    
  91.    [-1, 1, Conv, [128, 3, 2]],
  92.    [[-1, 47], 1, Concat, [1]],
  93.    
  94.    [-1, 1, Conv, [64, 1, 1]],
  95.    [-2, 1, Conv, [64, 1, 1]],
  96.    [-1, 1, Conv, [64, 3, 1]],
  97.    [-1, 1, Conv, [64, 3, 1]],
  98.    [[-1, -2, -3, -4], 1, Concat, [1]],
  99.    [-1, 1, Conv, [128, 1, 1]],  # 65
  100.    
  101.    [-1, 1, Conv, [256, 3, 2]],
  102.    [[-1, 37], 1, Concat, [1]],
  103.    
  104.    [-1, 1, Conv, [128, 1, 1]],
  105.    [-2, 1, Conv, [128, 1, 1]],
  106.    [-1, 1, Conv, [128, 3, 1]],
  107.    [-1, 1, Conv, [128, 3, 1]],
  108.    [[-1, -2, -3, -4], 1, Concat, [1]],
  109.    [-1, 1, Conv, [256, 1, 1]],  # 73
  110.  
  111.    [47, 1, Conv, [64, 3, 1]],
  112.    [57, 1, Conv, [128, 3, 1]],
  113.    [65, 1, Conv, [256, 3, 1]],
  114.    [73, 1, Conv, [512, 3, 1]],
  115.  
  116.    [[74,75,76,77], 1, Detect, [nc, anchors]],   # Detect(P3, P4, P5)
  117.   ]

近期有朋友问到对于改进yolov7网络结构后增加小目标检测层，如下所示，yaml文件所涉及的感兴趣的朋友可以关注私信我：

  1. # parameters
  2. nc: 3  # number of classes
  3. depth_multiple: 0.33  # model depth multiple
  4. width_multiple: 0.50  # layer channel multiple
  5.  
  6. # anchors
  7. anchors:
  8.   - [ 19,27,  44,40,  38,94 ]  # P3/8
  9.   - [ 96,68,  86,152,  180,137 ]  # P4/16
  10.   - [ 140,301,  303,264,  238,542 ]  # P5/32
  11.   - [ 436,615,  739,380,  925,792 ]  # P6/64
  12.  
  13. backbone:
  14.   # [from, number, module, args]
  15.   [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
  16.  
  17.     
  18.    [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
  19.        # [-1,1,CoordAtt,[128]],
  20.    [-1, 3, ELANB, [128]],
  21.    [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
  22.    [-1, 6, ELANB, [256]],
  23.    [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
  24.    [-1, 9, ELANB, [512]],
  25.    [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
  26.    [-1, 3, ELANB, [1024]],
  27.    [-1, 1, SPPCSPC, [512]],
  28.   ]
  29.  
  30. head:
  31.   [[-1, 1, SimConv, [256, 1, 1]],
  32.    [-1, 1, Transpose, [256]],
  33.    [[-1, 6], 1, Concat, [1]],
  34.    [-1, 12, RepBlock, [256]],
  35.  
  36.    [-1, 1, SimConv, [256, 1, 1]],
  37.    [-1, 1, Transpose, [256]],
  38.    [[-1, 4], 1, Concat, [1]],
  39.    [-1, 12, RepBlock, [256]],
  40.  
  41.    [-1, 1, SimConv, [128, 1, 1]],
  42.    [-1, 1, Transpose, [128]],
  43.    [[-1, 2], 1, Concat, [1]],
  44.    [-1, 12, RepBlock, [128]],   #out
  45.  
  46.    [-1, 1, SimConv, [128, 3, 2]],
  47.    [[-1, 18], 1, Concat, [1]],
  48.    [-1, 12, RepBlock, [256]],  # 20
  49.  
  50.    [-1, 1, SimConv, [128, 3, 2]],
  51.    [[-1, 14], 1, Concat, [1]],
  52.    [-1, 12, RepBlock, [256]],  # 20
  53.  
  54.    [-1, 1, SimConv, [256, 3, 2]],
  55.    [[-1, 10], 1, Concat, [1]],  # cat head P5
  56.    [-1, 12, RepBlock, [512]],  # 23
  57.  
  58.    [[21,24,27,30], 1, IDetect, [nc, anchors]],   # Detect(P3, P4, P5)
  59.   ]

四、YOLOv5改进方法

YOLOv5改进YOLOv5s.yaml，改进方法参考YOLOv7算法改进。

  1. backbone:
  2.   # [from, number, module, args]
  3.   [[-1, 1, Focus, [64, 3]],  # 0-P1/2
  4.    [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
  5.    [-1, 3, C3, [128]],
  6.    [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
  7.    [-1, 9, C3, [256]],
  8.    [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
  9.    [-1, 9, C3, [512]],
  10.    [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
  11.    [-1, 1, SPP, [1024, [5, 9, 13]]],
  12.    [-1, 3, C3, [1024, False]],  # 9
  13.   ]
  14.  
  15. # YOLOv5 head
  16. head:
  17.   [[-1, 1, Conv, [512, 1, 1]],#20*20
  18.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  19.    [[-1, 6], 1, Concat, [1]],  # cat backbone P4
  20.    [-1, 3, C3, [512, False]],  # 13
  21.  
  22.    [-1, 1, Conv, [512, 1, 1]], #40*40 14
  23.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  24.    [[-1, 4], 1, Concat, [1]],  # cat backbone P3   80*80
  25.    [-1, 3, C3, [512, False]],  # 17 (P3/8-small)  80*80
  26.  
  27.    [-1, 1, Conv, [256, 1, 1]],
  28.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  29.    [[-1, 2], 1, Concat, [1]],  # cat backbone P3
  30.    [-1, 3, C3, [256, False]],  # 17 (P3/8-small)21
  31.  
  32.    [-1, 1, Conv, [256, 3, 2]],
  33.    [[-1, 18], 1, Concat, [1]],  # cat head P4
  34.    [-1, 3, C3, [256, False]],  # 20 (P4/16-medium)24
  35.  
  36.    [-1, 1, Conv, [256, 3, 2]],  #22   80*80
  37.    [[-1, 14], 1, Concat, [1]], #23 80*80
  38.    [-1, 3, C3, [512, False]], #24 80*80
  39.  
  40.    [-1, 1, Conv, [512, 3, 2]],
  41.    [[-1, 10], 1, Concat, [1]],  # cat head P5
  42.    [-1, 3, C3, [1024, False]],  # 23 (P5/32-large)
  43.  
  44.    [[21, 24, 27,30], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  45.   ]

 添加小目标检测层后的模型图如下所示：

 最后，将train.py中改为本文的yaml文件即可，开始训练。

结 果：本人在多个数据集上做了大量实验，针对不同的数据集效果不同，同一个数据集的不同添加位置方法也是有差异，需要大家进行实验。有效果有提升的情况占大多数。

预告一下：下一篇内容分享损失函数的改进。有兴趣的朋友可以关注一下我，有问题可以留言或者私聊我哦

PS：增加检测层的方法不仅仅是适用改进YOLOv5，也可以改进其他的YOLO网络，比如YOLOv4、v3等。

最后，有改进相关问题欢迎关注私信我。

文章知识点与官方知识档案匹配，可进一步学习相关知识
算法技能树首页概览41800 人正在系统学习中
人工智能目标检测
QQ名片
【算法系列篇】滑动窗口
走进人工智能|深度学习 算法的创世纪