基于caffe的多标签端到端车牌识别

最近利用深度学习框架caffe做了一个端到端车牌识别的模型，所用的网络为AlexNet稍加了一些修改，由于数据集太少、网络结构不完善，效果并没有很好，但是一些基本的思路摸清楚了。

一、数据处理

所用的数据为EasyPR提供的数据集，共有1541张车牌图片，按照8：1：1的比例分为train、val和test三个数据集。（数据集太少，后期还会用GAN和真实车牌生成数据）

caffe自带的lmdb数据库没有存储多标签数据的功能，所以用hdf5对数据进行存储。车牌共有7个字符，第一个字符为汉字，第二个字符为字母，后面五个字符为数字和字母，所以将标签分为三类：31个汉字标签，24个字母标签，以及34个字母数字标签。

opencv读取的图片为H*W*C的形式，根据blob对图片reshape为C*H*W维度。

# 读取labelmap
labeldict_Chi = dict()
labelmap_Chi = open("labelmap_Chinese.txt", "r")
lines = labelmap_Chi.readlines()
for line in lines:
line = line.strip("\n")
temp = line.split(" ")
labeldict_Chi[temp[1]] = temp[0]
labelmap_Chi.close()

labeldict_Letter = dict()
labelmap_Letter = open("labelmap_Letter.txt", "r")
lines = labelmap_Letter.readlines()
for line in lines:
line = line.strip("\n")
temp = line.split(" ")
labeldict_Letter[temp[1]] = temp[0]
labelmap_Letter.close()

labeldict_NL = dict()
labelmap_NL = open("labelmap_Num&Letter.txt", "r")
lines = labelmap_NL.readlines()
for line in lines:
line = line.strip("\n")
temp = line.split(" ")
labeldict_NL[temp[1]] = temp[0]
labelmap_NL.close()

for i in range(2):
imgs_file = os.listdir(IMG_FILE[i])

for img_file in imgs_file:
filename = os.path.join(IMG_FILE[i], img_file)
image = cv2.imread(filename)
image = image.reshape(3, 36, 136)

label = []
char_Chi = img_file[:3]
label.append(int(labeldict_Chi[char_Chi]))
char_Spe = img_file[3:4]
label.append(int(labeldict_Letter[char_Spe]))
for char in img_file[4:-4]:
label.append(int(labeldict_NL[char]))

IMAGE[i].append(image)
LABEL[i].append(np.array(label))

IMAGE[i] = np.array(IMAGE[i], dtype=int)
LABEL[i] = np.array(LABEL[i], dtype=int)

meanValue = mean(IMAGE[0])

print "Creating hdf5..."

4000
with h5py.File(HDF5_FILE[i], "w") as h5:
data = IMAGE[i].astype(np.float32) - meanValue
label = LABEL[i].astype(np.float32)
h5["data"] = data
h5["label"] = label
h5.close()

print "Creating txt..."
with open(TXT_FILE[i], "w") as txt:
txt.write(os.path.join(os.getcwd(), HDF5_FILE[i]))
txt.close()

np.save("mean_e2e.npy", meanValue)

将图片存入h5文件的时候，需要加上图片的均值，由于caffe计算均值的文件是针对lmdb格式的，所以均值就自己写了一个：

# 计算均值
def mean(images):
images = images.astype(np.float32)
meanValue = sum(images) / len(images)
return meanValue

最后要将h5文件读入到一个txt文件里，这个txt文件存储的就是h5文件的路径。

二、生成网络结构

网络结构使用的是AlexNet结构，对网络结构进行了一些修改：

数据层用的是hdf5数据；

多标签分类，添加了slice层；

最后的输出根据标签数量的不同，输出的特征数有变化。

name: "Plate_E2E"
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
hdf5_data_param {
source: "examples/plate_e2e/train_e2e.txt"
batch_size: 10
}
}
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
include {
phase: TEST
}
hdf5_data_param {
source: "examples/plate_e2e/val_e2e.txt"
batch_size: 10
}
}
layer {
name: "slicers"
type: "Slice"
bottom: "label"
top: "label_1"
top: "label_2"
top: "label_3"
top: "label_4"
top: "label_5"
top: "label_6"
top: "label_7"
slice_param {
axis: 1
slice_point: 1
slice_point: 2
slice_point: 3
slice_point: 4
slice_point: 5
slice_point: 6
}
}
layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 96
kernel_size: 11
stride: 4
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu1"
type: "ReLU"
bottom: "conv1"
top: "conv1"
}
layer {
name: "norm1"
type: "LRN"
bottom: "conv1"
top: "norm1"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "pool1"
type: "Pooling"
bottom: "norm1"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "conv2"
type: "Convolution"
bottom: "pool1"
top: "conv2"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 256
pad: 2
kernel_size: 5
group: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu2"
type: "ReLU"
bottom: "conv2"
top: "conv2"
}
layer {
name: "norm2"
type: "LRN"
bottom: "conv2"
top: "norm2"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "pool2"
type: "Pooling"
bottom: "norm2"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "conv3"
type: "Convolution"
bottom: "pool2"
top: "conv3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 384
pad: 1
kernel_size: 3
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu3"
type: "ReLU"
bottom: "conv3"
top: "conv3"
}
layer {
name: "conv4"
type: "Convolution"
bottom: "conv3"
top: "conv4"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 384
pad: 1
kernel_size: 3
group: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu4"
type: "ReLU"
bottom: "conv4"
top: "conv4"
}
layer {
name: "conv5"
type: "Convolution"
bottom: "conv4"
top: "conv5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
group: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu5"
type: "ReLU"
bottom: "conv5"
top: "conv5"
}
layer {
name: "fc6"<
1271d
/span>
type: "InnerProduct"
bottom: "conv5"
top: "fc6"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu6"
type: "ReLU"
bottom: "fc6"
top: "fc6"
}
layer {
name: "drop6"
type: "Dropout"
bottom: "fc6"
top: "fc6"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7_1"
type: "InnerProduct"
bottom: "fc6"
top: "fc7_1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu7_1"
type: "ReLU"
bottom: "fc7_1"
top: "fc7_1"
}
layer {
name: "drop7_1"
type: "Dropout"
bottom: "fc7_1"
top: "fc7_1"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7_2"
type: "InnerProduct"
bottom: "fc6"
top: "fc7_2"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu7_2"
type: "ReLU"
bottom: "fc7_2"
top: "fc7_2"
}
layer {
name: "drop7_2"
type: "Dropout"
bottom: "fc7_2"
top: "fc7_2"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7_3"
type: "InnerProduct"
bottom: "fc6"
top: "fc7_3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu7_3"
type: "ReLU"
bottom: "fc7_3"
top: "fc7_3"
}
layer {
name: "drop7_3"
type: "Dropout"
bottom: "fc7_3"
top: "fc7_3"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7_4"
type: "InnerProduct"
bottom: "fc6"
top: "fc7_4"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu7_4"
type: "ReLU"
bottom: "fc7_4"
top: "fc7_4"
}
layer {
name: "drop7_4"
type: "Dropout"
bottom: "fc7_4"
top: "fc7_4"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7_5"
type: "InnerProduct"
bottom: "fc6"
top: "fc7_5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu7_5"
type: "ReLU"
bottom: "fc7_5"
top: "fc7_5"
}
layer {
name: "drop7_5"
type: "Dropout"
bottom: "fc7_5"
top: "fc7_5"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7_6"
type: "InnerProduct"
bottom: "fc6"
top: "fc7_6"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu7_6"
type: "ReLU"
bottom: "fc7_6"
top: "fc7_6"
}
layer {
name: "drop7_6"
type: "Dropout"
bottom: "fc7_6"
top: "fc7_6"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7_7"
type: "InnerProduct"
bottom: "fc6"
top: "fc7_7"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu7_7"
type: "ReLU"
bottom: "fc7_7"
top: "fc7_7"
}
layer {
name: "drop7_7"
type: "Dropout"
bottom: "fc7_7"
top: "fc7_7"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc8_1"
type: "InnerProduct"
bottom: "fc7_1"
top: "fc8_1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 31
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "fc8_2"
type: "InnerProduct"
bottom: "fc7_2"
top: "fc8_2"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 24
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "fc8_3"
type: "InnerProduct"
bottom: "fc7_3"
top: "fc8_3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 34
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "fc8_4"
type: "InnerProduct"
bottom: "fc7_4"
top: "fc8_4"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 34
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "fc8_5"
type: "InnerProduct"
bottom: "fc7_5"
top: "fc8_5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 34
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "fc8_6"
type: "InnerProduct"
bottom: "fc7_6"
top: "fc8_6"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 34
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "fc8_7"
type: "InnerProduct"
bottom: "fc7_7"
top: "fc8_7"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 34
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "accuracy_1"
type: "Accuracy"
bottom: "fc8_1"
bottom: "label_1"
top: "accuracy_1"
include {
phase: TEST
}
}
layer {
name: "accuracy_2"
type: "Accuracy"
bottom: "fc8_2"
bottom: "label_2"
top: "accuracy_2"
include {
phase: TEST
}
}
layer {
name: "accuracy_3"
type: "Accuracy"
bottom: "fc8_3"
bottom: "label_3"
top: "accuracy_3"
include {
phase: TEST
}
}
layer {
name: "accuracy_4"
type: "Accuracy"
bottom: "fc8_4"
bottom: "label_4"
top: "accuracy_4"
include {
phase: TEST
}
}
layer {
name: "accuracy_5"
type: "Accuracy"
bottom: "fc8_5"
bottom: "label_5"
top: "accuracy_5"
include {
phase: TEST
}
}
layer {
name: "accuracy_6"
type: "Accuracy"
bottom: "fc8_6"
bottom: "label_6"
top: "accuracy_6"
include {
phase: TEST
}
}
layer {
name: "accuracy_7"
type: "Accuracy"
bottom: "fc8_7"
bottom: "label_7"
top: "accuracy_7"
include {
phase: TEST
}
}
layer {
name: "loss_1"
type: "SoftmaxWithLoss"
bottom: "fc8_1"
bottom: "label_1"
top: "loss_1"
loss_weight: 0.2
}
layer {
name: "loss_2"
type: "SoftmaxWithLoss"
bottom: "fc8_2"
bottom: "label_2"
top: "loss_2"
loss_weight: 0.2
}
layer {
name: "loss_3"
type: "SoftmaxWithLoss"
bottom: "fc8_3"
bottom: "label_3"
top: "loss_3"
loss_weight: 0.2
}
layer {
name: "loss_4"
type: "SoftmaxWithLoss"
bottom: "fc8_4"
bottom: "label_4"
top: "loss_4"
loss_weight: 0.2
}
layer {
name: "loss_5"
type: "SoftmaxWithLoss"
bottom: "fc8_5"
bottom: "label_5"
top: "loss_5"
loss_weight: 0.2
}
layer {
name: "loss_6"
type: "SoftmaxWithLoss"
bottom: "fc8_6"
bottom: "label_6"
top: "loss_6"
loss_weight: 0.2
}
layer {
name: "loss_7"
type: "SoftmaxWithLoss"
bottom: "fc8_7"
bottom: "label_7"
top: "loss_7"
loss_weight: 0.2
}

有两个问题需要总结下：

AlexNet在conv5之后又一个pool5层，但是到conv5之后，图片的边界已经不足以再做池化，所以去掉了pool5层；

第一次训练的时候忘记修改最后fc层输出的num_output参数，导致特征分类出现问题。七个fc层的num_output分别是31、24、34、34、34、34、34，分别对应各自标签的类别。

solver修改为如下所示，再修改相应的deploy网络结构，就可以开始训练了。

net: "examples/plate_e2e/train_val_e2e.prototxt"
test_iter: 130
test_interval: 100
base_lr: 0.001
lr_policy: "step"
gamma: 0.1
stepsize: 3000
display: 10
max_iter: 15000
momentum: 0.9
weight_decay: 0.0005
snapshot: 3000
snapshot_prefix: "examples/plate_e2e/plate_e2e"
solver_mode: GPU

三、训练

#!/usr/bin/env sh

TOOLS=./build/tools

$TOOLS/caffe train --solver=examples/plate_e2e/solver_e2e.prototxt -gpu all

四、测试

def Test(img):
# 加载模型
net = caffe.Net(deploy, caffe_model, caffe.TEST)

# 注意可以调节预处理批次的大小
# 由于是处理一张图片，所以把原来的10张的批次改为1
net.blobs['data'].reshape(1, 3, 36, 136)

# 加载图片到数据层
im = cv2.imread(img)
images = np.array([im.reshape(3, 36, 136)], dtype=np.float32)
meanValue = mean()
net.blobs['data'].data[...] =  images - meanValue

# 前向计算
out = net.forward()

# 预测分类
result = []
result.append(out['prob_1'].argmax())
result.append(out['prob_2'].argmax())
result.append(out['prob_3'].argmax())
result.append(out['prob_4'].argmax())
result.append(out['prob_5'].argmax())
result.append(out['prob_6'].argmax())
result.append(out['prob_7'].argmax())

return result

最后的测试结果并不好，还需要对网络结构进行修改和构建，数据集也需要重新扩充，后期会进行更新。