368 lines
14 KiB
Python
368 lines
14 KiB
Python
# 导入必要的库
|
||
import torch
|
||
import torch.nn as nn
|
||
import cv2
|
||
import numpy as np
|
||
import os
|
||
import sys
|
||
from torch.autograd import Variable
|
||
from PIL import Image
|
||
|
||
# 添加父目录到路径，以便导入模型和数据加载器
|
||
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
||
|
||
# LPRNet字符集定义（与训练时保持一致）
|
||
# 包含中国省份简称、数字、字母和特殊字符
|
||
CHARS = ['京', '沪', '津', '渝', '冀', '晋', '蒙', '辽', '吉', '黑',
|
||
'苏', '浙', '皖', '闽', '赣', '鲁', '豫', '鄂', '湘', '粤',
|
||
'桂', '琼', '川', '贵', '云', '藏', '陕', '甘', '青', '宁', '新',
|
||
'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', 'I', 'O', '-']
|
||
|
||
# 创建字符到索引的映射字典
|
||
CHARS_DICT = {char: i for i, char in enumerate(CHARS)}
|
||
|
||
# 简化的LPRNet模型定义 - 基础卷积块
|
||
class small_basic_block(nn.Module):
|
||
def __init__(self, ch_in, ch_out):
|
||
super(small_basic_block, self).__init__()
|
||
# 定义一个小的基本卷积块，包含四个卷积层
|
||
self.block = nn.Sequential(
|
||
# 1x1卷积，降低通道数
|
||
nn.Conv2d(ch_in, ch_out // 4, kernel_size=1),
|
||
nn.ReLU(),
|
||
# 3x1卷积，处理水平特征
|
||
nn.Conv2d(ch_out // 4, ch_out // 4, kernel_size=(3, 1), padding=(1, 0)),
|
||
nn.ReLU(),
|
||
# 1x3卷积，处理垂直特征
|
||
nn.Conv2d(ch_out // 4, ch_out // 4, kernel_size=(1, 3), padding=(0, 1)),
|
||
nn.ReLU(),
|
||
# 1x1卷积，恢复通道数
|
||
nn.Conv2d(ch_out // 4, ch_out, kernel_size=1),
|
||
)
|
||
|
||
def forward(self, x):
|
||
return self.block(x)
|
||
|
||
# LPRNet模型定义 - 车牌识别网络
|
||
class LPRNet(nn.Module):
|
||
def __init__(self, lpr_max_len, phase, class_num, dropout_rate):
|
||
super(LPRNet, self).__init__()
|
||
self.phase = phase
|
||
self.lpr_max_len = lpr_max_len
|
||
self.class_num = class_num
|
||
|
||
# 定义主干网络
|
||
self.backbone = nn.Sequential(
|
||
# 初始卷积层
|
||
nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1), # 0
|
||
nn.BatchNorm2d(num_features=64),
|
||
nn.ReLU(), # 2
|
||
# 最大池化层
|
||
nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 1, 1)),
|
||
# 第一个基本块
|
||
small_basic_block(ch_in=64, ch_out=128), # *** 4 ***
|
||
nn.BatchNorm2d(num_features=128),
|
||
nn.ReLU(), # 6
|
||
# 第二个池化层
|
||
nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(2, 1, 2)),
|
||
# 第二个基本块
|
||
small_basic_block(ch_in=64, ch_out=256), # 8
|
||
nn.BatchNorm2d(num_features=256),
|
||
nn.ReLU(), # 10
|
||
# 第三个基本块
|
||
small_basic_block(ch_in=256, ch_out=256), # *** 11 ***
|
||
nn.BatchNorm2d(num_features=256),
|
||
nn.ReLU(), # 13
|
||
# 第三个池化层
|
||
nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(4, 1, 2)), # 14
|
||
# Dropout层，防止过拟合
|
||
nn.Dropout(dropout_rate),
|
||
# 特征提取卷积层
|
||
nn.Conv2d(in_channels=64, out_channels=256, kernel_size=(1, 4), stride=1), # 16
|
||
nn.BatchNorm2d(num_features=256),
|
||
nn.ReLU(), # 18
|
||
# 第二个Dropout层
|
||
nn.Dropout(dropout_rate),
|
||
# 分类卷积层
|
||
nn.Conv2d(in_channels=256, out_channels=class_num, kernel_size=(13, 1), stride=1), # 20
|
||
nn.BatchNorm2d(num_features=class_num),
|
||
nn.ReLU(), # 22
|
||
)
|
||
|
||
# 定义容器层，用于融合全局上下文信息
|
||
self.container = nn.Sequential(
|
||
nn.Conv2d(in_channels=448+self.class_num, out_channels=self.class_num, kernel_size=(1,1), stride=(1,1)),
|
||
)
|
||
|
||
def forward(self, x):
|
||
# 保存中间特征
|
||
keep_features = list()
|
||
for i, layer in enumerate(self.backbone.children()):
|
||
x = layer(x)
|
||
# 保存特定层的输出特征
|
||
if i in [2, 6, 13, 22]: # [2, 4, 8, 11, 22]
|
||
keep_features.append(x)
|
||
|
||
# 处理全局上下文信息
|
||
global_context = list()
|
||
for i, f in enumerate(keep_features):
|
||
# 对不同层的特征进行不同尺度的平均池化
|
||
if i in [0, 1]:
|
||
f = nn.AvgPool2d(kernel_size=5, stride=5)(f)
|
||
if i in [2]:
|
||
f = nn.AvgPool2d(kernel_size=(4, 10), stride=(4, 2))(f)
|
||
# 对特征进行归一化处理
|
||
f_pow = torch.pow(f, 2)
|
||
f_mean = torch.mean(f_pow)
|
||
f = torch.div(f, f_mean)
|
||
global_context.append(f)
|
||
|
||
# 拼接全局上下文特征
|
||
x = torch.cat(global_context, 1)
|
||
# 通过容器层处理
|
||
x = self.container(x)
|
||
# 对序列维度进行平均，得到最终输出
|
||
logits = torch.mean(x, dim=2)
|
||
|
||
return logits
|
||
|
||
# LPRNet推理类
|
||
class LPRNetInference:
|
||
def __init__(self, model_path=None, img_size=[94, 24], lpr_max_len=8, dropout_rate=0.5):
|
||
"""
|
||
初始化LPRNet推理类
|
||
Args:
|
||
model_path: 训练好的模型权重文件路径
|
||
img_size: 输入图像尺寸 [width, height]
|
||
lpr_max_len: 车牌最大长度
|
||
dropout_rate: dropout率
|
||
"""
|
||
self.img_size = img_size
|
||
self.lpr_max_len = lpr_max_len
|
||
# 检测是否有可用的CUDA设备
|
||
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
||
|
||
# 设置默认模型路径
|
||
if model_path is None:
|
||
current_dir = os.path.dirname(os.path.abspath(__file__))
|
||
model_path = os.path.join(current_dir, 'LPRNet__iteration_74000.pth')
|
||
|
||
# 初始化模型
|
||
self.model = LPRNet(lpr_max_len=lpr_max_len, phase=False, class_num=len(CHARS), dropout_rate=dropout_rate)
|
||
|
||
# 加载模型权重
|
||
if model_path and os.path.exists(model_path):
|
||
print(f"Loading LPRNet model from {model_path}")
|
||
try:
|
||
self.model.load_state_dict(torch.load(model_path, map_location=self.device))
|
||
print("LPRNet模型权重加载成功")
|
||
except Exception as e:
|
||
print(f"Warning: 加载模型权重失败: {e}. 使用随机权重.")
|
||
else:
|
||
print(f"Warning: 模型文件不存在或未指定: {model_path}. 使用随机权重.")
|
||
|
||
# 将模型移动到指定设备并设置为评估模式
|
||
self.model.to(self.device)
|
||
self.model.eval()
|
||
|
||
print(f"LPRNet模型加载完成，设备: {self.device}")
|
||
print(f"模型参数数量: {sum(p.numel() for p in self.model.parameters()):,}")
|
||
|
||
def preprocess_image(self, image_array):
|
||
"""
|
||
预处理图像数组 - 使用与训练时相同的预处理方式
|
||
Args:
|
||
image_array: numpy数组格式的图像 (H, W, C)
|
||
Returns:
|
||
preprocessed_image: 预处理后的图像tensor
|
||
"""
|
||
if image_array is None:
|
||
raise ValueError("Input image is None")
|
||
|
||
# 确保图像是numpy数组
|
||
if not isinstance(image_array, np.ndarray):
|
||
raise ValueError("Input must be numpy array")
|
||
|
||
# 检查图像维度
|
||
if len(image_array.shape) != 3:
|
||
raise ValueError(f"Expected 3D image array, got {len(image_array.shape)}D")
|
||
|
||
height, width, channels = image_array.shape
|
||
if channels != 3:
|
||
raise ValueError(f"Expected 3 channels, got {channels}")
|
||
|
||
# 调整图像尺寸到模型要求的尺寸
|
||
if height != self.img_size[1] or width != self.img_size[0]:
|
||
image_array = cv2.resize(image_array, tuple(self.img_size))
|
||
|
||
# 使用与训练时相同的预处理方式
|
||
# 归一化处理：减去127.5并乘以0.0078125，将像素值从[0,255]映射到[-1,1]
|
||
image_array = image_array.astype('float32')
|
||
image_array -= 127.5
|
||
image_array *= 0.0078125
|
||
# 调整维度顺序从HWC到CHW
|
||
image_array = np.transpose(image_array, (2, 0, 1)) # HWC -> CHW
|
||
|
||
# 转换为tensor并添加batch维度
|
||
image_tensor = torch.from_numpy(image_array).unsqueeze(0)
|
||
|
||
return image_tensor
|
||
|
||
def decode_prediction(self, logits):
|
||
"""
|
||
解码模型预测结果 - 使用正确的CTC贪婪解码
|
||
Args:
|
||
logits: 模型输出的logits [batch_size, num_classes, sequence_length]
|
||
Returns:
|
||
predicted_text: 预测的车牌号码
|
||
"""
|
||
# 转换为numpy进行处理
|
||
prebs = logits.cpu().detach().numpy()
|
||
preb = prebs[0, :, :] # 取第一个batch [num_classes, sequence_length]
|
||
|
||
# 贪婪解码： 对每个时间步选择最大概率的字符
|
||
preb_label = []
|
||
for j in range(preb.shape[1]): # 遍历每个时间步
|
||
preb_label.append(np.argmax(preb[:, j], axis=0))
|
||
|
||
# CTC解码：去除重复字符和空白字符
|
||
no_repeat_blank_label = []
|
||
pre_c = preb_label[0]
|
||
|
||
# 处理第一个字符
|
||
if pre_c != len(CHARS) - 1: # 不是空白字符
|
||
no_repeat_blank_label.append(pre_c)
|
||
|
||
# 处理后续字符
|
||
for c in preb_label:
|
||
if (pre_c == c) or (c == len(CHARS) - 1): # 重复字符或空白字符
|
||
if c == len(CHARS) - 1:
|
||
pre_c = c
|
||
continue
|
||
no_repeat_blank_label.append(c)
|
||
pre_c = c
|
||
|
||
# 转换为字符
|
||
decoded_chars = [CHARS[idx] for idx in no_repeat_blank_label]
|
||
return ''.join(decoded_chars)
|
||
|
||
def predict(self, image_array):
|
||
"""
|
||
预测单张图像的车牌号码
|
||
Args:
|
||
image_array: numpy数组格式的图像
|
||
Returns:
|
||
prediction: 预测的车牌号码
|
||
confidence: 预测置信度
|
||
"""
|
||
try:
|
||
# 预处理图像
|
||
image = self.preprocess_image(image_array)
|
||
if image is None:
|
||
return None, 0.0
|
||
|
||
image = image.to(self.device)
|
||
|
||
# 模型推理
|
||
with torch.no_grad():
|
||
logits = self.model(image)
|
||
# logits shape: [batch_size, class_num, sequence_length]
|
||
|
||
# 计算置信度（使用softmax后的最大概率平均值）
|
||
probs = torch.softmax(logits, dim=1)
|
||
max_probs = torch.max(probs, dim=1)[0]
|
||
confidence = torch.mean(max_probs).item()
|
||
|
||
# 解码预测结果
|
||
prediction = self.decode_prediction(logits)
|
||
|
||
return prediction, confidence
|
||
|
||
except Exception as e:
|
||
print(f"预测图像失败: {e}")
|
||
return None, 0.0
|
||
|
||
# 全局变量，用于存储模型实例
|
||
lpr_model = None
|
||
|
||
def LPRNinitialize_model():
|
||
"""
|
||
初始化LPRNet模型
|
||
|
||
返回:
|
||
bool: 初始化是否成功
|
||
"""
|
||
global lpr_model
|
||
|
||
try:
|
||
# 模型权重文件路径
|
||
model_path = os.path.join(os.path.dirname(__file__), 'LPRNet__iteration_74000.pth')
|
||
|
||
# 创建推理对象
|
||
lpr_model = LPRNetInference(model_path)
|
||
|
||
print("LPRNet模型初始化完成")
|
||
return True
|
||
|
||
except Exception as e:
|
||
print(f"LPRNet模型初始化失败: {e}")
|
||
import traceback
|
||
traceback.print_exc()
|
||
return False
|
||
|
||
def LPRNmodel_predict(image_array):
|
||
"""
|
||
LPRNet车牌号识别接口函数
|
||
|
||
参数:
|
||
image_array: numpy数组格式的车牌图像，已经过矫正处理
|
||
|
||
返回:
|
||
list: 包含最多8个字符的列表，代表车牌号的每个字符
|
||
例如: ['京', 'A', '1', '2', '3', '4', '5'] (蓝牌7位)
|
||
['京', 'A', 'D', '1', '2', '3', '4', '5'] (绿牌8位)
|
||
"""
|
||
global lpr_model
|
||
|
||
if lpr_model is None:
|
||
print("LPRNet模型未初始化，请先调用LPRNinitialize_model()")
|
||
return ['待', '识', '别', '0', '0', '0', '0', '0']
|
||
|
||
try:
|
||
# 使用OpenCV调整图像大小到模型要求的尺寸
|
||
image_array = cv2.resize(image_array, (128, 48))
|
||
print(f"666999图片尺寸: {image_array.shape}")
|
||
# 预测车牌号
|
||
predicted_text, confidence = lpr_model.predict(image_array)
|
||
|
||
if predicted_text is None:
|
||
print("LPRNet识别失败")
|
||
return ['识', '别', '失', '败', '0', '0', '0', '0']
|
||
|
||
print(f"LPRNet识别结果: {predicted_text}, 置信度: {confidence:.3f}")
|
||
|
||
# 将字符串转换为字符列表
|
||
char_list = list(predicted_text)
|
||
|
||
# 确保返回至少7个字符，最多8个字符
|
||
if len(char_list) < 7:
|
||
# 如果识别结果少于7个字符，用'0'补齐到7位
|
||
char_list.extend(['0'] * (7 - len(char_list)))
|
||
elif len(char_list) > 8:
|
||
# 如果识别结果多于8个字符，截取前8个
|
||
char_list = char_list[:8]
|
||
|
||
# 如果是7位，补齐到8位以保持接口一致性（第8位用空字符或占位符）
|
||
if len(char_list) == 7:
|
||
char_list.append('') # 添加空字符作为第8位占位符
|
||
|
||
return char_list
|
||
|
||
except Exception as e:
|
||
print(f"LPRNet识别失败: {e}")
|
||
import traceback
|
||
traceback.print_exc()
|
||
return ['识', '别', '失', '败', '0', '0', '0', '0'] |