深度学习和目标检测系列教程 17-300： 3 个类别面罩检测类别数据集训练yolov5s模型
Posted 刘润森！
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了深度学习和目标检测系列教程 17-300： 3 个类别面罩检测类别数据集训练yolov5s模型相关的知识，希望对你有一定的参考价值。
@Author：Runsen
YOLO 是目前最火爆的的计算机视觉算法之一，今天使用的数据集来源：https://www.kaggle.com/andrewmvd/face-mask-detection
这是数据集可以创建一个模型来检测戴口罩、不戴口罩或不正确戴口罩的人。
该数据集包含属于3 个类别的853张图像，以及它们的 PASCAL VOC 格式的边界框。分别是
- 带口罩；
- 不戴口罩；
- 口罩佩戴不正确。
首先将voc形式变成yolo形式，voc是xml文件，需要通过ElementTree进行解析，得到bnd_box ，具体解析代码如下
import os
from xml.etree.ElementTree import ElementTree
def xml_to_yolo(path):
root = ElementTree().parse(path)
img_path = root.find('filename').text.replace('png','txt')
img_size = root.find('size')
width = int(img_size.find('width').text)
height = int(img_size.find('height').text)
with open('../dataset/label/' + img_path, 'w') as f:
lines = []
for node in root.findall('object'):
object_ = dict(class_=None, x=None, y=None, width=None, height=None)
# class
class_name = node.find('name').text
if(class_name == 'without_mask'):
object_['class_'] = '0'
elif(class_name == 'with_mask'):
object_['class_'] = '1'
else:
object_['class_'] = '2'
# bounding box
bnd_box = node.find("bndbox")
x_min = float(bnd_box[0].text)
y_min = float(bnd_box[1].text)
x_max = float(bnd_box[2].text)
y_max = float(bnd_box[3].text)
dw = float(1/width)
dh = float(1/height)
w = float(x_max - x_min)
h = float(y_max - y_min)
x = float((x_min + x_max)/2 -1)
y = float((y_min + y_max)/2 -1)
w = float(w * dw)
h = float(h * dh)
x = float(x * dw)
y = float(y * dh)
object_['x'] = str(x)
object_['y'] = str(y)
object_['width'] = str(w)
object_['height'] = str(h)
line = object_['class_'] + ' ' + object_['x'] + ' ' + object_['y'] + ' ' + object_['width'] + ' ' + object_['height']
lines.append(line)
lines.append('\\n')
for line in lines[:-1]:
f.write(line)
f.close
def process_data():
img_paths = []
for dirname, _, filenames in os.walk('../dataset/images'):
for filename in filenames:
img_paths.append(os.path.join('../dataset/images', filename)) # local machine
# xml
for dirname, _, filenames in os.walk('../dataset/xml'):
for filename in filenames:
annotation_path = (os.path.join(dirname, filename))
xml_to_yolo(annotation_path)
process_data()
克隆Github的YOLOv5的代码：https://github.com/ultralytics/yolov5。
为了跟数据集的目录保持一致
在dataset新建images和labels文件夹存放对于的文件。
data文件夹下的data.yaml文件
train: ./dataset/images/train
val: ./dataset/images/val
# number of classes
nc: 3
# class names
names: ['without_mask','with_mask','mask_weared_incorrect']
model文件夹下的yolov5s.yaml文件
# Parameters
nc: 3 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple
anchors:
- [10,13, 16,30, 33,23] # P3/8
- [30,61, 62,45, 59,119] # P4/16
- [116,90, 156,198, 373,326] # P5/32
# YOLOv5 backbone
backbone:
# [from, number, module, args]
[[-1, 1, Focus, [64, 3]], # 0-P1/2
[-1, 1, Conv, [128, 3, 2]], # 1-P2/4
[-1, 3, C3, [128]],
[-1, 1, Conv, [256, 3, 2]], # 3-P3/8
[-1, 9, C3, [256]],
[-1, 1, Conv, [512, 3, 2]], # 5-P4/16
[-1, 9, C3, [512]],
[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
[-1, 1, SPP, [1024, [5, 9, 13]]],
[-1, 3, C3, [1024, False]], # 9
]
# YOLOv5 head
head:
[[-1, 1, Conv, [512, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 6], 1, Concat, [1]], # cat backbone P4
[-1, 3, C3, [512, False]], # 13
[-1, 1, Conv, [256, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 4], 1, Concat, [1]], # cat backbone P3
[-1, 3, C3, [256, False]], # 17 (P3/8-small)
[-1, 1, Conv, [256, 3, 2]],
[[-1, 14], 1, Concat, [1]], # cat head P4
[-1, 3, C3, [512, False]], # 20 (P4/16-medium)
[-1, 1, Conv, [512, 3, 2]],
[[-1, 10], 1, Concat, [1]], # cat head P5
[-1, 3, C3, [1024, False]], # 23 (P5/32-large)
[[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5)
]
下载yolov5s.pt对应的权重，大约14M
训练代码：python train.py --batch 16 --epochs 50 --data /data/data.yaml --cfg /models/yolov5s.yaml
训练结果
关于P R mAP50 和mAP95的指标
mAP50，mAP60，mAP70……等等指的是取detector的IoU阈值大于0.5，大于0.6，大于0.7…的比例。数值越高，越好，如果阈值越大，精度会变低。
因为IoU大于0.95的小于IoU大于0.5
yolov5训练完，会生成runs的文件夹
训练结束会生成train文件夹，第一次训练完成一般默认是exp，如果再训练一次，就是exp2。exp19是我在不断的改batch 和epochs。
50次训练的指标结果
epoch, train/box_loss, train/obj_loss, train/cls_loss, metrics/precision, metrics/recall, metrics/mAP_0.5,metrics/mAP_0.5:0.95, val/box_loss, val/obj_loss, val/cls_loss, x/lr0, x/lr1, x/lr2
0, 0.11183, 0.061692, 0.036166, 0.0040797, 0.023207, 0.00094919, 0.00015348, 0.088802, 0.044058, 0.029341, 0.00053, 0.00053, 0.09523
1, 0.090971, 0.06624, 0.024651, 0.44734, 0.13502, 0.11474, 0.0279, 0.067798, 0.037852, 0.022663, 0.0010692, 0.0010692, 0.090369
2, 0.070342, 0.052357, 0.02153, 0.85191, 0.20848, 0.22242, 0.077028, 0.055487, 0.027934, 0.019595, 0.0016049, 0.0016049, 0.085505
3, 0.0699, 0.043899, 0.020244, 0.87295, 0.16442, 0.20657, 0.078501, 0.055767, 0.024396, 0.018845, 0.0021348, 0.0021348, 0.080635
4, 0.064317, 0.04001, 0.018747, 0.83973, 0.19818, 0.20992, 0.068715, 0.055507, 0.022381, 0.018095, 0.0026562, 0.0026562, 0.075756
5, 0.060517, 0.040462, 0.01871, 0.92677, 0.26977, 0.31691, 0.13554, 0.046723, 0.020782, 0.017638, 0.0031668, 0.0031668, 0.070867
6, 0.055565, 0.037719, 0.017239, 0.56328, 0.42257, 0.34731, 0.14537, 0.042455, 0.020308, 0.015934, 0.0036641, 0.0036641, 0.065964
7, 0.051137, 0.036675, 0.01443, 0.70063, 0.59002, 0.56658, 0.28446, 0.037626, 0.018761, 0.012734, 0.0041459, 0.0041459, 0.061046
8, 0.051564, 0.036086, 0.010754, 0.75331, 0.59469, 0.57015, 0.21859, 0.044253, 0.018153, 0.010504, 0.00461, 0.00461, 0.05611
9, 0.054336, 0.036489, 0.01014, 0.81292, 0.53958, 0.57799, 0.28552, 0.037579, 0.018368, 0.009468, 0.0050544, 0.0050544, 0.051154
10, 0.052323, 0.034572, 0.0095445, 0.85048, 0.62312, 0.6267, 0.338, 0.031516, 0.017655, 0.010057, 0.005477, 0.005477, 0.046177
11, 0.050382, 0.035035, 0.0086806, 0.86829, 0.59491, 0.61092, 0.3379, 0.032883, 0.017608, 0.0090742, 0.0058761, 0.0058761, 0.041176
12, 0.046103, 0.035051, 0.0082707, 0.8746, 0.60331, 0.72019, 0.42562, 0.030186, 0.016849, 0.0087845, 0.00625, 0.00625, 0.03615
13, 0.045144, 0.032905, 0.0071727, 0.50942, 0.80319, 0.76972, 0.38975, 0.033539, 0.016185, 0.0083475, 0.0065973, 0.0065973, 0.031097
14, 0.043005, 0.031431, 0.0076075, 0.94694, 0.59886, 0.75893, 0.45433, 0.028811, 0.016204, 0.0087185, 0.0069167, 0.0069167, 0.026017
15, 0.042583, 0.032493, 0.0075345, 0.54628, 0.78986, 0.70483, 0.39052, 0.034077, 0.016426, 0.010784, 0.007207, 0.007207, 0.020907
16, 0.044892, 0.032174, 0.0079312, 0.93707, 0.62972, 0.75324, 0.417, 0.031565, 0.015583, 0.0074945, 0.0074674, 0.0074674, 0.015767
17, 0.04351, 0.033322, 0.0071257, 0.9655, 0.60346, 0.75329, 0.43919, 0.029984, 0.015935, 0.0071412, 0.0076971, 0.0076971, 0.010597
18, 0.041603, 0.032276, 0.0070445, 0.56572, 0.84931, 0.80283, 0.50439, 0.027445, 0.015567, 0.0067956, 0.0077031, 0.0077031, 0.0077031
19, 0.03878, 0.031363, 0.0064713, 0.57055, 0.85292, 0.81709, 0.54818, 0.02845, 0.014806, 0.006212以上是关于深度学习和目标检测系列教程 17-300： 3 个类别面罩检测类别数据集训练yolov5s模型的主要内容，如果未能解决你的问题，请参考以下文章
深度学习和目标检测系列教程 3-300：了解常见的目标检测的开源数据集
深度学习和目标检测系列教程 19-300：关于目标检测APIoU和mAP简介
深度学习和目标检测系列教程 19-300：关于目标检测APIoU和mAP简介
深度学习和目标检测系列教程 14-300：训练第一个 YOLOv3 检测器