pytorch图像识别（基于识别验证码），读取数据集图片以及使用神经网络进行训练

数据集介绍

数据集代码可以看博客：验证码图片生成
图像数据集为10000张验证码图片，图片每次为验证码内容加数字（后面加入0-10000的数字是为了防止出现一样的验证码）：

数据集读取

读取文件使用os模块：

from os import listdir

pics = listdir('./data')
for i in pics:
    print(i)

接下来需要加上路径，这样一会处理的话就可以直接读取每张图片：

dirs = []
pics = listdir('./data')
for i in pics:
    path = './data/' + i
    dirs.append(path)
print(dirs)

图片读取功能完成了。

使用onehot编码

首先验证码范围为（0-9，a-z，A-Z），这些装入列表中：

captcha_array = list(string.ascii_letters + string.digits)

图片类型为./data/00Ei_2196.png，需要获取其中的00Ei，我将使用正则表达式：

import re

name = re.search(r'data/(.*?)_', dirs[0])
print(name.group(1))

后续为转为数字（就是每个符号在符号列表的位置）：

for i in name.group(1):
    x.append(captcha_array.index(i))
print(x)

转为one hot编码：

print(x)
x = torch.Tensor(x)
print(x)
y = F.one_hot(x.to(torch.int64), len(captcha_array))
print(y)
print(y.shape)

数据集读取和数据转为one hot封装：

dataset.py

import torch
import torch.nn.functional as F
import string
from os import listdir
import re
import os
from PIL import Image
from torch.utils.data import Dataset
from torchvision import transforms


class my_dataset(Dataset):
    def __init__(self, dir):
        super(my_dataset, self).__init__()
        pics = listdir('./data')
        self.paths = [os.path.join(dir, image_name) for image_name in pics]
        self.transforms = transforms.Compose(
            [
                transforms.ToTensor(),  # 转为tensor类型
            ]

        )
        self.captcha_array = list(string.ascii_letters + string.digits)
        self.url = dir

    def __len__(self):
        return self.paths.__len__()

    def __getitem__(self, index):
        path = self.paths[index]
        image = Image.open(path).convert("RGB")
        data = self.transforms(image)
        label = re.search(rf'{self.url}(.*?)_', path).group(1)
        x = list()
        for i in label:
            x.append(self.captcha_array.index(i))
        x = torch.Tensor(x)
        label = F.one_hot(x.to(torch.int64), len(self.captcha_array))

        return data, label



if __name__ == '__main__':
    train_data = my_dataset("./data/")
    data, label = train_data[8]
    print(data, label)
    print(data.shape)
    print(label.shape)

编写神经网络

model_one.py

import torch
from torch import nn
from dataset import my_dataset
from torch.utils.data import DataLoader
from torch import nn
from torch.optim import Adam
import time
import tqdm
class mymodel(nn.Module):
    def __init__(self):
        super(mymodel, self).__init__()
        self.layer1=nn.Sequential(
            nn.Conv2d(in_channels=3,out_channels=64,kernel_size=3,stride=1,padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)#[64, 64, 30, 90]
        )
        self.layer2=nn.Sequential(
            nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2)
        )

        self.layer3 = nn.Sequential(
          nn.Flatten(),#[6, 2560] [64, 15360]
          nn.Linear(in_features=86400,out_features=4096),
          nn.Dropout(0.2),  # drop 20% of the neuron
          nn.ReLU(),
          nn.Linear(in_features=4096, out_features=62 * 4)
        )
    def forward(self,x):
        x=self.layer1(x)
        #print('layer1:',x.shape)
        x=self.layer2(x)
        #print('layer2:', x.shape)
        x=self.layer3(x)
        #print('layer3:', x.shape)
        x = x.view(x.size(0), -1)
        return x;

训练：

import torch
from torch import nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

import my_datasets
from model import mymodel

if __name__ == '__main__':
    train_datas= my_datasets.mydatasets("./dataset/train")
    test_data= my_datasets.mydatasets("./dataset/test")
    train_dataloader=DataLoader(train_datas,batch_size=64,shuffle=True)
    test_dataloader = DataLoader(test_data, batch_size=64, shuffle=True)
    writer=SummaryWriter("logs")
    m=mymodel().cuda()

    loss_fn=nn.MultiLabelSoftMarginLoss().cuda()
    optimizer = torch.optim.Adam(m.parameters(), lr=0.001)
    w=SummaryWriter("logs")
    total_step=0

for i in range(10):
    for i,(imgs,targets) in enumerate(train_dataloader):
        imgs=imgs.cuda()
        targets=targets.cuda()
        # print(imgs.shape)
        # print(targets.shape)
        outputs=m(imgs)
        # print(outputs.shape)
        loss = loss_fn(outputs, targets)
        optimizer.zero_grad()

        loss.backward()
        optimizer.step()

    writer.close()
torch.save(m,"model.pth")

测试

我的测试集是5000张验证码图片

from tqdm import tqdm
from dataset import my_dataset
from torch.utils.data import DataLoader
import torch
import string

captcha_array = list(string.ascii_letters + string.digits)
def Text(vec):
    """
    将onehot编码转为标注文本
    :return text
    """
    text = ""
    for i in vec:
        text += captcha_array[i]
    return text

if __name__ == '__main__':
    test_dataset = my_dataset("./ce/")
    # print(test_dataset[0])
    test_dataloader = DataLoader(test_dataset, batch_size=1, shuffle=True)
    # 定义网络
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
#cnn.pth Verification_resnet34_4w.pth
    model = torch.load("cnn.pth").cuda(device)
    total = len(test_dataset)
    yuce = 0
    bar = tqdm(enumerate(test_dataloader), total=len(test_dataloader))
    for i, (images, labels) in bar:
        images = images.cuda(0)
        labels = labels.cuda(0)
        outputs = model(images)
        # outputs = outputs.reshape(1, 4, 62)
        outputs = outputs.view(-1, 62)
        labels = labels.view(-1, 62)
        pred = torch.max(outputs, 1)[1].cpu().detach().numpy()
        labels = torch.max(labels, 1)[1].cpu().detach().numpy()
        pre = Text(pred)
        label = Text(labels)

        if pre == label:
            yuce += 1
            print(pre, label)
        else:
            # print(pred, labels)
            pass
    print("正确率: {}%".format(yuce / total * 100))

准确率为99.82%，可以效果非常好了。

同时还有一个方法，深度残差网络（ResNet），我是直接调用库里的：

import torch
import time
from tqdm import tqdm
from dataset import my_dataset
from torch.utils.data import DataLoader
from torch import nn
from torch.optim import Adam
import torchvision.models as models


if __name__ == '__main__':
    train_dataset  = my_dataset("./data/")
    train_dataloader = DataLoader(train_dataset, batch_size=64, shuffle=True, drop_last=True, num_workers=4)
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

    verification = models.resnet34(num_classes=4 * 62).cuda(device)
    loss_fn = nn.MSELoss().cuda(device)
    optim = Adam(verification.parameters(), lr=0.001)
    verification.train()
    for epoch in range(50):
        now = time.time()
        bar = tqdm(enumerate(train_dataloader), total=len(train_dataloader))

        for i, (images, labels) in bar:
            optim.zero_grad()
            images = images.cuda(device)
            labels = labels.cuda(device)
            outputs = verification(images)
            outputs = outputs.reshape(64, 4, 62)
            labels = labels.to(torch.float)
            loss = loss_fn(outputs, labels)
            loss.backward()
            optim.step()
        torch.save(verification, "resnet34_4w.pth")
        print("训练epoch次数{}".format(epoch + 1))

准确率也是极高的：