--- a detection target moving bricks ALPR automatic number plate recognition environment
reference
License Plate Detection and Recognition in Unconstrained Scenarios
@https: //www.cnblogs.com/greentomlee/p/10863363.html
@https: //github.com/ sergiomsilva / alpr-unconstrained
environment
the current version was tested in an Ubuntu 16.04 machine, with Keras 2.2.4, TensorFlow 1.5.0, OpenCV 2.4.9, NumPy 1.14 and Python 2.7.
versions OpenCV difference and evolution, from 1.x to 4.0
@https: //www.cnblogs.com/shine-lee/p/9884551.html
build a virtual environment isolated python, different versions of opencv: Ubuntu 16.04 + Keras 2.2.4 + TensorFlow 1.5.0 + openCV 2.4. . 9 + NumPy Python 1.14 + 2.7
(1) activates the virtual environment
u @ u160406: ~ / Virtualenv / py2 $ source bin / activate
(py2) u160406 U @: ~ / Virtualenv / py2 $
(2) OpenCV-Python
automatically install OpenCV
Python -m PIP install Python OpenCV-
PIP-install OpenCV Python
PIP install OpenCV-contrib-Python
install the latest version of opencv ( here Successfully installed opencv-python-4.1.1.26)
view version Python
Import CV2
CV2 .__ version__
>>> 4.1.1
no opencv 2.x version of
python -m pip install opencv-python == 2.4.9
ERROR: Could not find a version that satisfies the requirement opencv-python == 2.4.9 (from versions: 3.1.0.0, 3.1.0.1, 3.1.0.2, 3.1.0.3, 3.1.0.4, 3.1.0.5, 3.2.0.6 , 3.2.0.7, 3.2.0.8, 3.3.0.9, 3.3.0.10 , 3.3.1.11, 3.4.0.12, 3.4.0.14, 3.4.1.15, 3.4.2.16, 3.4.2.17, 3.4.3.18, 3.4.4.19, 3.4 .5.20, 3.4.6.27, 3.4.7.28, 4.0.0.21, 4.0.1.23, 4.0.1.24, 4.1.0.25, 4.1.1.26)
unloading
PIP Python the install OpenCV-
PIP-uninstall OpenCV Python-4.1.1.26-contrib
(2 ') manually install OpenCV
. 1, download
wget -O opencv.zip https://github.com/Itseez/opencv/archive/2.4.9.zip
wget -O opencv-3.3.0.zip https://github.com /opencv/opencv/archive/2.4.9.zip
contrib repository: HTTPS: //github.com/opencv/opencv_contrib/releases
OpenCV version: https: //opencv.org/releases.html
opencv2.x comes contrib repository, contrib repository from opencv 3.x + version has been independent, so opencv2.x without downloading opencv_contrib.zip, download can not find resources.
opencv2.x mounted without separate installation contrib library, opencv 3.x + contrib need to install a separate library
https://github.com/Itseez/opencv_contrib/archive/2.4.9.zip wget -O opencv_contrib.zip
wget -O opencv- https://github.com/opencv/opencv_contrib/archive/2.4.9.zip 3.3.0.zip
2, OpenCV into the folder, create a build directory, compile:
cd OpenCV-2.4.9
mkdir build
3, into the build directory:
cd build
OpenCV 2.x seemingly does not support CUDA, cmake has been an error, and finally through the following command:
cmake options:
cmake .. -DCMAKE_BUILD_TYPE Release -DCUDA_nppi_LIBRARY = = = OFF to true -DWITH_CUDA ..
cmake -D CMAKE_BUILD_TYPE = Release -D CMAKE_INSTALL_PREFIX = / home / u / Virtualenv / py2 / local / -D PYTHON_EXECUTABLE = / home / u / Virtualenv / py2 / bin / python -D PYTHON_PACKAGES_PATH = / home / u / Virtualenv / py2 / lib / python2.7 / Site-Packages ..
. 4, the compiler starts: the make -j4
. 5, the installation: the make the install the sudo
. 6, test whether the installation is successful
Python
>>> Import CV2
>>> version__ CV2 .__
'2.4.9'
( 3) exit the virtual environment
(py2) u160406 U @: ~ / Virtualenv / py2 / deactivate cars-LPR $
(4) because in a virtual environment are installed opencv, because there is no configuration in the system opencv environment variable.
Restart the computer to see if opencv version can also use
U @ u160406: ~ / Virtualenv / Source py2 $ bin / of an activate
(py2) u160406 U @: ~ / Virtualenv / py2 $ Python
Python 2.7.12 (default, Nov 12 2018, 14: 36:49)
[20,160,609 the GCC 5.4.0] ON linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import cv2
>>> cv2.__version__
'2.4.9'
>>> 可用
*************************************************************************************************************************************************
@https://blog.csdn.net/ELEVEN_ZOU/article/details/80893579
@https: //www.cnblogs.com/greentomlee/p/10863363.html
A, WHAT
Paper Download: License Plate Detection and Recognition in Unconstrained Scenarios [ pdf ]
GitHub project Address: ALPR-unconstrained
Data set: http://www.inf.ufrgs.br/~crjung/alpr-datasets.
Project Home: ALPR-Datasets
Video Effects: Demi Lovato Rock in Riio Lisboa 2018
Selected herein ECCV2018 paper " License Plate Detection and Recognition in Unconstrained Scenarios (complex | license plate detection and recognition under unconstrained scenes)." The paper does not give into the car with a complete identification system (Automatic License Plate Recognition system, ALPR system) solutions, but also provides recognition algorithm in the scene in unconstrained (Unconstrained Scenarios), a good solution license plate recognition problems in real life.
1. The introduction of novel CNN, the license plate can be detected and corrected more distortions of the single image, after the correct license plate images into ocr method to obtain a final text of the license plate
2. On the license plate from different regions were manually labeled: Use a really trained by the license plate data and the composition of artificially generated data, and uses a font similar to the target area re-training class network ocr
3. Expand the existing OCR methods were developed for the Brazilian license plates.
Results: The accuracy rate is high not only in the Brazilian license plates, license plates born in Europe and Taiwan also gained a high recall and precision
PS: Taiwan and Europe license plate license plate is composed of an array of letters and, in fact, constitute Brazil and the license plate is the same, the only difference is that the proportion of all license plate image may vary. But this image preprocessing when images like a unified set
二、WHY
The most common commercial software license plate identification, monitoring, such as car parks, toll booths in the case of the scene is limited, these scenes are fixed, but the camera angles and license plate for loose scenes, such as the hands of law enforcement officers the camera is a smart machine or captured pictures, photos inclination may be high, for these commercial software is concerned, it is difficult to identify them.
Therefore, this paper presents a complete ALPR system that can perform well under varying scenarios.
In general is this: At present, many ALPR system for the front license plate photos taken good recognition results, but in the changing angles and scene under (light) license plates are often tilted, resulting in the identification of effects and unhappy. Therefore, this paper presents a complete ALPR system, can be varied scene perform well under.
Data set exists inclined plate
Based on this, this article presents a complete ALPR system, can have a good performance in a variety of scenarios. The main contribution of this paper is twofold:
1. describes a novel neural network, the license plate can be detected under different environmental scenarios; OCR and before deformation is corrected .
2. Use real data synthesis of a variety of license plate, used for training. Manually marked sample of less than 200, and is a training network from scratch .
Three, HOW (implementation)
The main task is to find and identify ALPR license (license plates) in the image; which would normally be divided into four sub-task sequence:
- Detecting vehicle (vehicle detection)
- License Plate Detection (license plate detection)
- Character segmentation (character segmentation)
- Character recognition (character recognition)
The sub-tasks 3 and 4 can be seen as OCR; so a total of three sub-tasks: Vehicle Detection, License Plate Detection, OCR article also conduct carried out under the order of the three sub-tasks. As shown below .
Specifically, it is: YOLOv3 vehicle identification based on - the OCR license plate recognition> -> detection and correction of plates:
Examples of processes
Question 1: Why is first detected in the vehicle, and then detect the license plate? Rather than direct detection of the license plate it?
根据经验，较大的输入图像上是可以检测较小的物体，但是这会增加了计算量。车牌属于小目标的检测，通常情况下在有车的图片上，往往会伴随着车牌，（即：有车牌的地方一定能检测出车辆），所以可以先检测车辆。
问题2： 为什么对车牌进行校正？
这是因为在正视图上，车牌的大小和车辆的bounding box 的比例较大；而在斜视图或者侧视图上，车牌大小和车辆的比例较小。因此，需要将倾斜的视图调整到正面的视图上，以保持车牌区域的可识别性。 简而言之就是：倾斜视角的车牌面积较小，将车牌变成正视图的视角后，车牌面积会增大，从而增加车牌的识别率。
倾斜视角的车牌面积较小，将车牌变成正视图的视角后，车牌面积会增大，从而增加车牌的识别率。
问题3： 车牌的校正是如何实现的？
在车牌校正方面，虽然3D姿态估计是方法可以确定校正的尺寸和调整的尺度，但是本文提出了一种基于车辆bounding box的纵横比（aspect ratio）的一种方法，简单又高效。首先需要设置resizing factor,即当bounding box的aspect ratio接近于1的时候，影响因子会较小；当aspect ratio 增加的时候，影响因子也会增加。
缩放因子的计算方式如下：
（1）中，Wv和 Hv是识别车辆后的图像大小， Dmin和Dmax是常量，分别是288和608
其中：Wv和Hv是车辆的bounding box的宽度和高度。
注意： 由于Dmin ≤ fsc min(Wv, Hv) ≤ Dmax，所以Dmin和Dmax为调整车辆图片的维度大小划定了范围，经过尝试之后最佳的值是Dmin=288； Dmax=608
fsc代码实现如下：
1 Ivehicle = cv2.imread(img_path) # vehicle bounding box
2
3 ratio = float(max(Ivehicle.shape[:2]))/min(Ivehicle.shape[:2]) # 计算缩放比例 4 side = int(ratio*288.) 5 bound_dim = min(side + (side%(2**4)),608) # fsc 6 print "\t\tBound dim: %d, ratio: %f" % (bound_dim,ratio) 7 8 factor = float(max_dim)/min_dim_img 9 print "\t Fsc is %f"%(factor)
缩放因子用于计算后期的图像的宽高：
1 # wpod_net前期对车辆的bounding box 的归一化
2 # 详情见：def detect_lp(model,I,max_dim,net_step,out_size,threshold):
3 def detect_lp(model,I,max_dim,net_step,out_size,threshold): 4 ''' 5 :param model: 车牌检测模型，wpod_net 6 :param I: 车辆图片，im2single(Ivehicle)，W,H,C 7 :param max_dim: bound_dim 8 :param net_step:2**4 9 :param out_size:(240,80) 10 :param threshold:lp_threshold==0.5 11 :return: 12 ''' 13 14 min_dim_img = min(I.shape[:2])# [W,H,C]-->[W,H] 15 factor =.....# 计算的缩放因子 16 17 w,h = (np.array(I.shape[1::-1],dtype=float)*factor).astype(int).tolist()# [w,h,c] 18 w += (w%net_step!=0)*(net_step - w%net_step) 19 h += (h%net_step!=0)*(net_step - h%net_step) 20 Iresized = cv2.resize(I,(w,h))# 车辆图片做缩放 21 22 T = Iresized.copy() 23 T = T.reshape((1,T.shape[0],T.shape[1],T.shape[2])) 24 25 start = time.time() # 获取时间 26 Yr = model.predict(T) # 预测的车牌区域 27 Yr = np.squeeze(Yr) 28 elapsed = time.time() - start 29 ....(略)#后期是放射变换，校正车牌的 30 31
3.1车辆的检测
车辆检测所需要的数据集是：Cars Dataset
考虑车辆的检测的需要使用的较少的时间，于是借用了YUOLOv3的模型与darknet框架进行检测的；
使用 YOLOv3的主要原因是：
1 . 检测速度很快（大约是70 FPS ）;
2. 识别的准确率较高，在PASCAL VOC数据集中的测试结果是76.8% mPA.
在这篇文章中，将YOLOv3直接拿来使用，并没有做性能上的改进，同时忽略了除了cars以外的其他类别；
PS :
Yolov3 相比于Faster RCNN 确实很快，但是也有致命的缺陷： 不能识别出特别小的物体。如果在实际应用中需要在一张很大的图像上识别很小的车，建议重新训练
另外，YOLOv2 中具有车牌的物体不仅仅是cars， 还有bus和truck；这两个类别在程序中确被忽略的。
使用YOLOv2 用于车辆检测：本文没有对YOLO进行更改和改进，只是将网络该网络作为车辆检测的工具，将汽车和公交车这两类进行合并，忽略了yolo里面的其他类别。
将检测出的车辆图片裁剪并且进行调整，作为车牌识别的输入。
测出的正样本（车辆），将会被resized ，送入车牌检测模块。resize 因子已经在上一节给出，通过这种缩放计算以后，包含有车辆的图片会被统一成大小为 288*608大小的图片；
3.2 车牌的检测和校正（License Plate Detection and Unwarping）
车牌检测所需的数据集是 SSIG License Plate Character Segmentation Database 和 AOLPR
车牌本质上是附加在车上的矩形平面物体。为了充分的利用车牌的形状优势，本文提出了一种 扭曲平面检测网络（Warped Planar Object Detection Network., WPOD NET）。该网络可以学习识别不同扭曲程度的车牌，并学习回归仿射变换的系数，将就扭曲的车牌unwarps成正面视图下的矩形形状。WOPD的检测流程如下图所示。
经过WPOD网络之后是一个8通道的特征图（图片中显示的是6通道，应该是8）
在特征图上提取扭曲的车牌，首先会设置固定大小的单元（m, n）
如果该单元的目标概率大于给定的阈值（threshold），那么部分回归参数用于构建放射变换矩阵，将虚拟正方形区域转换成车牌的区域
该网络的结构总共21层卷积，其中14层是内部残差块（ResBlock）；所有的卷积核大小均为固定大小的3*3；除了识别块（Detection）以外，其他的激励函数全部是ReLu；包含有4个2*2大小的max pooling，stride=2，这样可以使得输入的维度减少了16倍（2*2*4）；最后在识别模块有两个并行的卷积层
具体结构如下图（fig4）所示，左侧是网络的整体结构，右侧是ResBlock和Detection块的结构。
可以看出来，在检测出车牌以后，还会对倾斜的车牌进行校正；现在需要关注的部分有两点：
1. WPOD是如何设计的？
2. 如何进行车牌校正的？
3. 为什么会在最后一层使用两个并行的卷积层？
检测车牌的整个流程
问题1. WPOD是如何设计的？
好，回答第一个问题： WPOD的设计：
网络结构如下，可以看出：整个网络是 Conv（ 包含res-block ） +Maxpooling +DETECTION构成的； 一共有21个卷积层，其中14个是residual blocks，所有的卷积层的filter/kernal大小都是3*3；并且每个Conv操作都会使用到ReLu; 另外有4个Maxpooling的，kernel=2*2, stride=2,
DETECTION上是两个并行的网络（粉红色的区域）; 第一个是使用softmax函数去推断概率； 第二个是回归仿射参数，而不需要激活，使用恒等式F(X)=x作为激活函数
WPOD net 结构详情
1 # 为了书写方便，将conv+bn+relue封装在一个函数里面
2 def conv_batch(_input,fsz,csz,activation='relu',padding='same',strides=(1,1)):
3 output = Conv2D(fsz, csz, activation='linear', padding=padding, strides=strides)(_input)
4 output = BatchNormalization()(output) 5 output = Activation(activation)(output) 6 ## 最后两层的书写方式： 7 def end_block(x): 8 xprobs = Conv2D(2, 3, activation='softmax', padding='same')(x) 9 xbbox = Conv2D(6, 3, activation='linear' , padding='same')(x) 10 return Concatenate(3)([xprobs,xbbox]) 11 12 ## model的构建 13 def create_model_eccv(): 14 15 input_layer = Input(shape=(None,None,3),name='input') 16 17 x = conv_batch(input_layer, 16, 3) 18 x = conv_batch(x, 16, 3) 19 x = MaxPooling2D(pool_size=(2,2))(x) 20 x = conv_batch(x, 32, 3) 21 x = res_block(x, 32) 22 x = MaxPooling2D(pool_size=(2,2))(x) 23 x = conv_batch(x, 64, 3) 24 x = res_block(x,64) 25 x = res_block(x,64) 26 x = MaxPooling2D(pool_size=(2,2))(x) 27 x = conv_batch(x, 64, 3) 28 x = res_block(x,64) 29 x = res_block(x,64) 30 x = MaxPooling2D(pool_size=(2,2))(x) 31 x = conv_batch(x, 128, 3) 32 x = res_block(x,128) 33 x = res_block(x,128) 34 x = res_block(x,128) 35 x = res_block(x,128) 36 37 x = end_block(x) 38 39 return Model(inputs=input_layer,outputs=x) 40 ##
res_block的实现：
1 def res_block(x,sz,filter_sz=3,in_conv_size=1):
2 xi = x
3 for i in range(in_conv_size):
4 xi = Conv2D(sz, filter_sz, activation='linear', padding='same')(xi)
5 xi = BatchNormalization()(xi) 6 xi = Activation('relu')(xi) 7 xi = Conv2D(sz, filter_sz, activation='linear', padding='same')(xi) 8 xi = BatchNormalization()(xi)# xi 的通道和x的通道是一样的 9 xi = Add()([xi,x]) ## 这里是矩阵的加法 10 xi = Activation('relu')(xi) 11 return xi
问题2. 车牌区域是如何校正的？
如下图所示，WPOD Net本身是没有纠正车牌的能力；
作者是构建仿射矩阵，将虚拟方框（白色的方框）转换为lp区域。 WPOD net会输出一个8-channel的feature map,以及编码对象的概率和放射变换的参数。为了能够获取扭曲的车牌，首先在box的中心设置一个固定大小的虚框。如果以（m,n）单元为中心的虚框的目标概率大于检测的阈值，说明白色方框覆盖的区域里 有车牌，同时为扭曲的车牌进行仿射变换。
1 # inference
2 Ivehicle = cv2.imread(img_path) # vehicle bounding box; 读取读取车辆图片（根据bounding box裁剪）
3 factor= ...# 计算缩放因子 4 w,h = ... # 根据缩放因子和车辆图像计算w和h 5 Iresized = cv2.resize(I,(w,h))# 车辆图片做缩放 6 7 T = T.reshape((1,T.shape[0],T.shape[1],T.shape[2])) 8 Yr= model.predict(T) # 预测的车牌区域 9 # reconstruct 10 Yr=Y 11 Probs = Y[...,0] # 模型的预测结果的概率 12 Affines = Y[...,2:] # 仿射变换的 13 xx,yy = np.where(Probs>threshold)#threshold=0.5 14 15 base = lambda vx,vy: np.matrix([[-vx,-vy,1.],[vx,-vy,1.],[vx,vy,1.],[-vx,vy,1.]]).T # 构建放射矩阵 16 17 for i in range(len(xx)): 18 y,x = xx[i],yy[i] 19 affine = Affines[y,x] 20 prob = Probs[y,x] 21 22 mn = np.array([float(x) + .5,float(y) + .5]) 23 24 A = np.reshape(affine,(2,3)) 25 A[0,0] = max(A[0,0],0.) 26 A[1,1] = max(A[1,1],0.) 27 28 pts = np.array(A*base(vxx,vyy)) #*alpha 29 pts_MN_center_mn = pts*side 30 pts_MN = pts_MN_center_mn + mn.reshape((2,1)) 31 32 pts_prop = pts_MN/MN.reshape((2,1)) 33 34 labels.append(DLabel(0,pts_prop,prob)) 35 36 ... #仿射变换end 37
对扭曲车牌检测的效果图：
问题3. 为什么会在最后一层使用两个并行的卷积层？
到现在为止，我们还没有回答这个网络的最关键的问题：文中提到的warp net放射变换参数是从哪里得到的？
网络最后的并行卷积层有着不同的使命：
第一个卷积模块是：用于计算物体的概率的，使用了softmax作为激励函数。
第二个卷积模块是：用于获取仿射参数的，没有使用激励函数
两个卷积学习的目标是根据Loss function 而定的，因此还需要知道loss function 是如何实现的。
第一个卷积模块是：用于计算物体的概率的，使用了softmax作为激励函数。
第二个卷积模块是：获取仿射参数的，没有使用激励函数
Loss Function 的设计：
假设
问题4. 如何实现仿射变换的？仿射变换恢复的扭曲图片和3D恢复的扭曲图片有什么区别？
1.解释仿射变换
2. 解释仿射变换实现的车牌校正
3. 解释仿射变换与3D扭曲图片之间的区别
3.3 车牌的OCR识别
这部分字符分割和识别是使用改进的YOLO模型。
所使用的改进方法和网络架构和这篇文章一样：Silva, S.M., Jung, C.R.: Real-time brazilian license plate detection and recognition using deep convolutional neural networks. In: 2017 30th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI). pp. 55–62 (Oct 2017). https://doi.org/10.1109/SIBGRAPI.2017.14
数据集是通过使用合成和扩增数据来处理世界各地不同区域的LP特征，在该工作中大大地扩大了训练数据集。 人工创建的数据包括将7个字符的字符串粘贴到纹理背景上，然后执行随机转换，如旋转、平移、噪声和模糊。
车牌数据集的制作流程如下图所示：
制作好的数据如下入所示：
这种综合数据的使用极大地提高了网络的泛化能力，使得完全相同的网络对世界不同地区的LPs都有很好的应用效果。
这部分涉及OCR , 后期会专门设置一个专题阐述OCR的识别算法
四、总结
这篇论文的实用性应该较高；但就创新性而言，并没有太多改进。值得一提的是对扭曲的车牌进行了仿射变换，使得车牌可以恢复正常矩形的形状 。
在论文的最后，作者给出测试的结果，通过对比5中不同的车牌检测数据集可以看出这篇论文的综合测试的正确率较高。
五、附录
5.1 检测车辆
1 1 mkdir tmp # 创建临时存储的文件夹
2 2 # 检测车辆，第一个参数是检测图片所在的目录， 第二个参数将检测的车辆保存的目录
3 3 python vehicle-detection.py ./samples/test/ ./tmp
5.2 检测车牌
1 python license-plate-detection.py ./tmp/ ./data/lp-detector/wpod-net_update1.h5
输出的日志：
1 Searching for license plates using WPOD-NET
2 Processing ./tmp/03033_0car.png
3 Bound dim: 476, ratio: 1.606780
4 Processing ./tmp/03016_0car.png
5 Bound dim: 606, ratio: 2.052117 6 Processing ./tmp/03071_0car.png 7 Bound dim: 608, ratio: 2.156250 8 Processing ./tmp/03057_0car.png 9 Bound dim: 526, ratio: 1.804819 10 Processing ./tmp/03066_0car.png 11 Bound dim: 544, ratio: 1.889503 12 Processing ./tmp/03058_0car.png 13 Bound dim: 346, ratio: 1.159664 14 Processing ./tmp/03066_1car.png 15 Bound dim: 490, ratio: 1.685950 16 Processing ./tmp/03066_2car.png 17 Bound dim: 608, ratio: 2.135135 18 Processing ./tmp/03025_0car.png 19 Bound dim: 568, ratio: 1.958333 20 Processing ./tmp/03058_1car.png 21 Bound dim: 372, ratio: 1.288136 22 Processing ./tmp/03009_0car.png 23 Bound dim: 526, ratio: 1.802211 24 Processing ./tmp/03009_1car.png 25 Bound dim: 608, ratio: 3.037433
5.3 ocr识别
1 python license-plate-ocr.py ./tmp
2
3 # log
4 layer filters size input output
5 0 conv 32 3 x 3 / 1 240 x 80 x 3 -> 240 x 80 x 32 0.033 BFLOPs 6 1 max 2 x 2 / 2 240 x 80 x 32 -> 120 x 40 x 32 7 2 conv 64 3 x 3 / 1 120 x 40 x 32 -> 120 x 40 x 64 0.177 BFLOPs 8 3 max 2 x 2 / 2 120 x 40 x 64 -> 60 x 20 x 64 9 4 conv 128 3 x 3 / 1 60 x 20 x 64 -> 60 x 20 x 128 0.177 BFLOPs 10 5 conv 64 1 x 1 / 1 60 x 20 x 128 -> 60 x 20 x 64 0.020 BFLOPs 11 6 conv 128 3 x 3 / 1 60 x 20 x 64 -> 60 x 20 x 128 0.177 BFLOPs 12 7 max 2 x 2 / 2 60 x 20 x 128 -> 30 x 10 x 128 13 8 conv 256 3 x 3 / 1 30 x 10 x 128 -> 30 x 10 x 256 0.177 BFLOPs 14 9 conv 128 1 x 1 / 1 30 x 10 x 256 -> 30 x 10 x 128 0.020 BFLOPs 15 10 conv 256 3 x 3 / 1 30 x 10 x 128 -> 30 x 10 x 256 0.177 BFLOPs 16 11 conv 512 3 x 3 / 1 30 x 10 x 256 -> 30 x 10 x 512 0.708 BFLOPs 17 12 conv 256 3 x 3 / 1 30 x 10 x 512 -> 30 x 10 x 256 0.708 BFLOPs 18 13 conv 512 3 x 3 / 1 30 x 10 x 256 -> 30 x 10 x 512 0.708 BFLOPs 19 14 conv 80 1 x 1 / 1 30 x 10 x 512 -> 30 x 10 x 80 0.025 BFLOPs 20 15 detection 21 mask_scale: Using default '1.000000' 22 Loading weights from data/ocr/ocr-net.weights...Done! 23 Performing OCR... 24 Scanning ./tmp/03009_0car_lp.png 25 LP: GN06BG 26 Scanning ./tmp/03016_0car_lp.png 27 LP: MPE3389 28 Scanning ./tmp/03025_0car_lp.png 29 LP: INS6012 30 Scanning ./tmp/03033_0car_lp.png 31 LP: SEZ229 32 Scanning ./tmp/03057_0car_lp.png 33 LP: INTT263 34 Scanning ./tmp/03058_1car_lp.png 35 LP: C24JBH 36 Scanning ./tmp/03066_0car_lp.png 37 LP: 77 38 Scanning ./tmp/03066_2car_lp.png 39 LP: HHP8586 40 Scanning ./tmp/03071_0car_lp.png 41 LP: 6GQR959
最终的识别结果：
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------
1.基本功能：从一张或者一系列的图片中提取车牌信息，比如车牌号码、车牌颜色等信息。
2.功能扩展：车型、车品牌、车牌类型等。
3.应用方向：电子交易系统（停车自动收费、收费站自动支付等）、交通执法、交通监控等。
4.影响因子:
(1)车牌的影响：位置、数量、尺寸、颜色、字体、背景图案和颜色（指车牌的背景）、遮挡（可能车牌较脏）、倾斜、特殊车牌、双行车牌、其他（比如车框、螺钉）。
(2)环境影响:拍摄相机的种类(彩色、黑白、红外）和拍摄角度、光照影响、背景（指车的外形背景）。
5.车牌识别的一般步骤：图象获取——>车牌提取——>车牌字符分割——>字符识别。
6.各种识别方法之间性能比较方案：比较方法的优缺点（应用范围、功能范围）、识别率、识别速度等。
7.双行车牌的定位与字符的切割。
8.中国车牌的特点：多省份多种类的中文字符，且中文字符可能出现的位置有三处，如下所示：
且还有一些受环境影响的车牌:
# 总结如下：
车牌字符：数字0--9（10个）、字母A--Z（没有I和O，24个）、省份简称（31个）、特殊字符（7个，还有一些比较少见，没算在内） 共计72个
"0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "A", "B", "C", "D", "E", "F", "G", "H", "J", "K", "L", "M", "N", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z","川","鄂","赣","甘","贵","桂","黑","泸","冀","津","京","吉","辽","鲁","蒙","闽","宁","青","琼",
"陕","苏","晋","皖","湘","新","豫","渝","粤","云","藏","浙","澳","港","挂","警","领","使","学"
环境因素：拍摄角度、分辨率、环境亮度、模糊程度等。
原文链接：https://blog.csdn.net/ELEVEN_ZOU/article/details/80893579