All Projects → Feeyao → License-plate-recognition

Feeyao / License-plate-recognition

Licence: MIT license
使用 "Darknet yolov3-tiny" 进行车牌识别

Programming Languages

C++
36643 projects - #6 most used programming language

Projects that are alternatives of or similar to License-plate-recognition

go-darknet
Go bindings for Darknet (YOLO v4 / v3)
Stars: ✭ 56 (-37.78%)
Mutual labels:  darknet, yolov3-tiny
darknet.js
A NodeJS wrapper of pjreddie's darknet / yolo.
Stars: ✭ 61 (-32.22%)
Mutual labels:  darknet, yolov3-tiny
Listen Attend Spell
A PyTorch implementation of Listen, Attend and Spell (LAS), an End-to-End ASR framework.
Stars: ✭ 147 (+63.33%)
Mutual labels:  end-to-end
DIoU YOLO V3
📈📈📈【口罩佩戴检测数据训练 | 开源口罩检测数据集和预训练模型】Train D/CIoU_YOLO_V3 by darknet for object detection
Stars: ✭ 53 (-41.11%)
Mutual labels:  darknet
quickstart-examples
Integration examples of Tanker's client-side encryption SDKs
Stars: ✭ 17 (-81.11%)
Mutual labels:  end-to-end
End2end Asr Pytorch
End-to-End Automatic Speech Recognition on PyTorch
Stars: ✭ 175 (+94.44%)
Mutual labels:  end-to-end
morghulis
No description or website provided.
Stars: ✭ 18 (-80%)
Mutual labels:  darknet
Rnn Transducer
MXNet implementation of RNN Transducer (Graves 2012): Sequence Transduction with Recurrent Neural Networks
Stars: ✭ 114 (+26.67%)
Mutual labels:  end-to-end
darknet
php ffi darknet
Stars: ✭ 21 (-76.67%)
Mutual labels:  darknet
Automatic speech recognition
End-to-end Automatic Speech Recognition for Madarian and English in Tensorflow
Stars: ✭ 2,751 (+2956.67%)
Mutual labels:  end-to-end
DarkPlate
License plate parsing using Darknet and YOLO
Stars: ✭ 36 (-60%)
Mutual labels:  darknet
Sstd
Single Shot Text Detector with Regional Attention
Stars: ✭ 221 (+145.56%)
Mutual labels:  end-to-end
Gun
An open source cybersecurity protocol for syncing decentralized graph data.
Stars: ✭ 15,172 (+16757.78%)
Mutual labels:  end-to-end
yolor
implementation of paper - You Only Learn One Representation: Unified Network for Multiple Tasks (https://arxiv.org/abs/2105.04206)
Stars: ✭ 1,867 (+1974.44%)
Mutual labels:  darknet
Eend
End-to-End Neural Diarization
Stars: ✭ 153 (+70%)
Mutual labels:  end-to-end
gravity
User-space deniable data encryption client.
Stars: ✭ 89 (-1.11%)
Mutual labels:  end-to-end
Deepvoice3 pytorch
PyTorch implementation of convolutional neural networks-based text-to-speech synthesis models
Stars: ✭ 1,654 (+1737.78%)
Mutual labels:  end-to-end
My bibliography for research on autonomous driving
Personal notes about scientific and research works on "Decision-Making for Autonomous Driving"
Stars: ✭ 197 (+118.89%)
Mutual labels:  end-to-end
darknet-nnpack
Darknet with NNPACK
Stars: ✭ 302 (+235.56%)
Mutual labels:  darknet
OpenCV-Flask
🐛 🐛 Opencv视频流传输到网页浏览器并做目标检测 🐛 🐛
Stars: ✭ 35 (-61.11%)
Mutual labels:  darknet

License-plate-recognition

使用 "Darknet yolov3-tiny" 训练检测模型

1. 下载data.zip,提取码: j7c2.

2. 将data.zip解压到darknet.exe所在目录下.

3. 进入data/voc目录下运行voc_label.bat重新生成2019_train.txt, 2019_val.txt.

4. 修改cfg/yolov3-tiny.cfg

[net]

batch=64

subdivisions=4    // 这里根据自己内存大小修改(我11G显存设置2时,中途会out of memory. 所以设置4, 训练时显存占用约6G)

angle=5           // 增加旋转角度产生样本

max_batches = 220000        //最大迭代次数

steps=70000,200000         //调整学习率变化点

...

filters=225                 //[yolo]前一个filters=(classes类别数+ coords坐标数 +1) * mask个数

[yolo]

anchors = 12,27,  17,45,  23,61,  37,58,  198,140,  344,319

classes=70

ignore_thresh = .7

...

其他参数说明可参考: https://blog.csdn.net/weixin_42731241/article/details/81474920

5. 执行

darknet.exe detector train data/voc.data cfg/yolov3-tiny.cfg

6. 训练过程(以其中一次过程为例)

训练过程也一直在调参数. 所以中断了几次.主要调节anchors 和 ignore_thresh

A. iter_0-34900 

      第一个 anchors=12,27,  15,34,  17,45,  23,61,  30,84,  198,140
      
      第二个 anchors=10,14,  23,27,  37,58,  81,82,  135,169,  344,319
      
      ignore_thresh=.8
      
B. iter_34900-42700

      两个 anchors=10,14,  23,27,  37,58,  81,82,  135,169,  344,319
      
C. iter_42700-65000

      ignore_thresh=.7
      
D. iter_65000-74700

      ignore_thresh=0.05
      
E. iter_74700-

      ignore_thresh=.7
      
F. 最终因为loss趋于平缓, 选择了结束训练. 当然还可以再降低learning_rate继续训练, 我这里没继续了.

loss

7. 测试了Test目录下的73张图片, 最终识别对了70张, 识别率约95.89%.weights, 提取码:rlh7

test

[总结]

  1. 此方法对输入图片存在一定要求, 车牌区域在图片上较小时, 字符可能检测不出或漏检. 所以测试时为输入改为608是提升字符检出率,不过识别时间约增加一倍. 更好的解决方法是不改变输入尺寸, 将车牌区域检测和字符检测分开两个模型, 先检测车牌区域, 将车牌区域(范围可稍微扩张一些)再进行字符检测.

  2. 训练数据里面倾斜样本和模糊样本少了一些, 对角度大且车牌区域小的图片字符检测效果不太好. image1image2image3

  3. 想起来再写.

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].