GEE_1.png (140.78 KB, 下载次数: 1)
下载附件
2025-8-31 22:34 上传
对于这种验证码, 第一眼看是不是毫无办法? 使用一般的轮廓识别/相似度比较已经不奏效了, 那怎么办呢? 经过长时间的搜索(Github), 我找到(并非自创)了这样一种解决办法(后续代码有一部分借鉴https://github.com/taisuii/ClassificationCaptchaOcr):
[ol]
[/ol]
准备数据集
下载
数据集我们需要约500张图(分割后的大图与小图标)下载到一个文件夹中, 这里就不说怎么获取图了, 只说一下怎么把九宫格图片分割成九张小图
from io import BytesIO
from PIL import Image
# 切割顺序，这里是从左到右，从上到下[x,y]
coordinates = [[1, 1], [1, 2], [1, 3], [2, 1], [2, 2], [2, 3], [3, 1], [3, 2], [3, 3]]
def crop_image(image_bytes, coordinates):
img = Image.open(BytesIO(image_bytes))
width, height = img.size
grid_width = width // 3
grid_height = height // 3
cropped_images = []
for coord in coordinates:
y, x = coord
left = (x - 1) * grid_width
upper = (y - 1) * grid_height
right = left + grid_width
lower = upper + grid_height
box = (left, upper, right, lower)
cropped_img = img.crop(box)
cropped_images.append(cropped_img)
return cropped_images
还需要对下载的小图进行去重 (部分下载的图片有重复)
import os
import hashlib
def file_md5(file_path):
"""计算文件的 MD5 值"""
hash_md5 = hashlib.md5()
with open(file_path, "rb") as f:
for chunk in iter(lambda: f.read(4096), b""):
hash_md5.update(chunk)
return hash_md5.hexdigest()
def deduplicate_by_content(folder_path):
seen = {}
for root, dirs, files in os.walk(folder_path):
for file in files:
file_path = os.path.join(root, file)
file_hash = file_md5(file_path)
if file_hash in seen:
print(f"删除重复文件: {file_path}")
os.remove(file_path)
else:
seen[file_hash] = file_path
# 对xxx文件夹下的文件进行去重
deduplicate_by_content(r"xxx")
标注
接下来新建一个文件夹, 这个文件夹应该为如下目录结构
datasets (当前文件夹)
├── images (此文件夹存储小图标)
├── outputs (此文件夹用于存放标注好的图标, 完成后内部会有很多文件夹)
├── Tool.py (我制作的标注工具, 一会给出代码)
其中Tool.py的内容如下:
import tkinter as tk
import os
from PIL import Image, ImageTk
def main():
# 创建Tkinter窗口
root = tk.Tk()
root.title("显示图片")
PC = [0]
Ls = os.listdir('images')
# 加载图片
image1 = Image.open('images/' + Ls[PC[0]])
image = ImageTk.PhotoImage(image1)
# 创建Label小部件并显示图片
label = tk.Label(image=image)
label.pack()
input_label = tk.Label(root, text="Input")
input_label.pack()
entry = tk.Entry(root)
entry.pack()
e = tk.Label(root, text="")
e.pack()
# submit
def submit(*args):
nonlocal image1
P = entry.get()
entry.delete(0, tk.END)
if P:
pass
else:
return
if os.path.exists('outputs/' + P):
pass
else:
os.mkdir('outputs/' + P)
image1.save('outputs/' + P + '/' + Ls[PC[0]])
PC[0] += 1
image1 = Image.open('images/' + Ls[PC[0]])
image = ImageTk.PhotoImage(image1)
label.config(image=image)
label.image = image
e.config(text=Ls[PC[0]])
submit_button = tk.Button(root, text="提交", command=submit) # 创建按钮，点击时调用get_input函数
entry.bind("", submit)
submit_button.pack()
# 进入Tkinter事件循环
root.mainloop()
if __name__ == "__main__":
main()
这个工具内部有一个输入框和一个按钮, 显示图片后你需要给同类图片给打上一样的标签, 然后按下回车或者按钮就会自动保存并接着下一张, 在全部标注完, 没有剩余图片时会报错
训练
使用Python安装上torch, torchvision, tqdm, numpy, onnxruntime, 这里的安装过程不再赘述, 在dateset文件夹的同级目录创建一个py文件和一个model文件夹, 写入如下内容并执行:
import torchvision.transforms as transforms
from torchvision.datasets import ImageFolder
from tqdm import tqdm
import torch
import torchvision
import torch.nn as nn
from torch.utils.data import DataLoader
import numpy as np
# 定义数据转换
data_transform = transforms.Compose(
[
transforms.Resize((224, 224)), # 调整图像大小
transforms.ToTensor(), # 将图像转换为张量
transforms.Normalize(
(0.485, 0.456, 0.406), (0.229, 0.224, 0.225)
), # 标准化图像
]
)
# 定义数据集
class CustomDataset:
def __init__(self, data_dir):
self.dataset = ImageFolder(root=data_dir, transform=data_transform)
def __len__(self):
return len(self.dataset)
def __getitem__(self, idx):
image, label = self.dataset[idx]
return image, label
class MyResNet18(torch.nn.Module):
def __init__(self, num_classes):
super(MyResNet18, self).__init__()
self.resnet = torchvision.models.resnet18(pretrained=True)
self.resnet.fc = nn.Linear(512, num_classes) # 修改这里的输入大小为512
def forward(self, x):
return self.resnet(x)
def train(epoch):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
data_dir = "dataset"
# 自定义数据集实例
custom_dataset = CustomDataset(data_dir)
# 数据加载器
batch_size = 64
data_loader = DataLoader(custom_dataset, batch_size=batch_size, shuffle=True)
# 初始化模型 num_classes就是目录下的子文件夹数目，每个子文件夹对应一个分类，模型输出的向量长度也是这个长度
model = MyResNet18(num_classes=91)
model.to(device)
# 损失函数
criterion = torch.nn.CrossEntropyLoss()
# 优化器
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
# 训练模型
for i in range(epoch):
losses = []
# 迭代器进度条
data_loader_tqdm = tqdm(data_loader)
for inputs, labels in data_loader_tqdm:
# 将输入数据和标签传输到指定的计算设备（如 GPU 或 CPU）
inputs, labels = inputs.to(device), labels.to(device)
# 梯度更新之前将所有模型参数的梯度置为零，防止梯度累积
optimizer.zero_grad()
# 前向传播：将输入数据传入模型，计算输出
outputs = model(inputs)
# 根据模型的输出和实际标签计算损失值
loss = criterion(outputs, labels)
# 将当前批次的损失值记录到 losses 列表中，以便后续计算平均损失
losses.append(loss.item())
epoch_loss = np.mean(losses)
data_loader_tqdm.set_description(
f"This epoch is {i} and it's loss is {loss.item()}, average loss {epoch_loss}"
)
# 反向传播：根据当前损失值计算模型参数的梯度
loss.backward()
# 使用优化器更新模型参数，根据梯度调整模型参数
optimizer.step()
# 每过一个batch就保存一次模型
torch.save(model.state_dict(), f'model/my_resnet18_{epoch_loss}.pth')
print(f"completed. Model saved.")
if __name__ == '__main__':
train(50)
待提示completed. Model saved.时即为训练成功, model里最近的模型文件就是训练好的模型
pth转为onnx
执行如下代码
from resnet18 import MyResNet18
import torch
def convert():
# 加载 PyTorch 模型
model_path = "model/改成你模型的文件名.pth"
model = MyResNet18(num_classes=91)
model.load_state_dict(torch.load(model_path))
model.eval()
# 生成一个示例输入
dummy_input = torch.randn(10, 3, 224, 224)
# 将模型转换为 ONNX 格式
torch.onnx.export(model, dummy_input, "model/resnet18.onnx", verbose=True)
if __name__ == '__main__':
convert()
使用
from PIL import Image
from io import BytesIO
import onnxruntime as ort
import numpy as np
# Resnet
class Resnet:
def __init__(self, ModelPath: str):
self.Session = ort.InferenceSession(ModelPath)
self.InputName = self.Session.get_inputs()[0].name
def cosine_similarity(self, vec1, vec2):
# 将输入转换为 NumPy 数组
vec1 = np.array(vec1)
vec2 = np.array(vec2)
# 计算点积
dot_product = np.dot(vec1, vec2)
# 计算向量的范数
norm_vec1 = np.linalg.norm(vec1)
norm_vec2 = np.linalg.norm(vec2)
# 计算余弦相似度
similarity = dot_product / (norm_vec1 * norm_vec2)
return similarity
def data_transforms(self, image):
image = image.resize((224, 224))
image_array = np.array(image)
image_array = image_array.astype(np.float32) / 255.0
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
image_array = (image_array - mean) / std
image_array = np.transpose(image_array, (2, 0, 1))
return image_array
# 识别
def classification(self, BG: bytes, Icon: bytes):
coordinates = [
[1, 1],
[1, 2],
[1, 3],
[2, 1],
[2, 2],
[2, 3],
[3, 1],
[3, 2],
[3, 3],
]
Icon = self.convert_png_to_jpg(Icon)
target_images = []
target_images.append(self.data_transforms(Image.open(BytesIO(Icon))))
bg_images = self.crop_image(BG, coordinates)
for bg_image in bg_images:
target_images.append(self.data_transforms(bg_image))
start = time.time()
outputs = self.Session.run(None, {self.InputName: target_images})[0]
scores = []
for i, out_put in enumerate(outputs):
if i == 0:
target_output = out_put
else:
similarity = self.cosine_similarity(target_output, out_put)
scores.append(similarity)
# 从左到右，从上到下，依次为每张图片的置信度
# print(scores)
# 对数组进行排序，保持下标
indexed_arr = list(enumerate(scores))
sorted_arr = sorted(indexed_arr, key=lambda x: x[1], reverse=True)
# 提取最大三个数及其下标
largest_three = sorted_arr[:3]
answer = [coordinates[i[0]] for i in largest_three]
return answer
def convert_png_to_jpg(self, png_bytes: bytes) -> bytes:
# 将传入的 bytes 转换为图像对象
png_image = Image.open(BytesIO(png_bytes))
# 创建一个 BytesIO 对象，用于存储输出的 JPG 数据
output_bytes = BytesIO()
# 检查图像是否具有透明度通道 (RGBA)
if png_image.mode == 'RGBA':
# 创建白色背景
white_bg = Image.new("RGB", png_image.size, (255, 255, 255))
# 将 PNG 图像粘贴到白色背景上，透明部分用白色填充
white_bg.paste(png_image, (0, 0), png_image)
jpg_image = white_bg
else:
# 如果图像没有透明度，直接转换为 RGB 模式
jpg_image = png_image.convert("RGB")
# 将转换后的图像保存为 JPG 格式到 BytesIO 对象
jpg_image.save(output_bytes, format="JPEG")
# 返回保存后的 JPG 图像的 bytes
return output_bytes.getvalue()
def crop_image(self, image_bytes, coordinates):
img = Image.open(BytesIO(image_bytes))
width, height = img.size
grid_width = width // 3
grid_height = height // 3
cropped_images = []
for coord in coordinates:
y, x = coord
left = (x - 1) * grid_width
upper = (y - 1) * grid_height
right = left + grid_width
lower = upper + grid_height
box = (left, upper, right, lower)
cropped_img = img.crop(box)
cropped_images.append(cropped_img)
return cropped_images
ResnetONNX = Resnet('我是模型.onnx')
ResnetONNX.classification(b'九宫格Bytes', b'小图标Bytes') # 返回结果即为图片位置, 例如: [[1, 2], [1, 3], [3, 1]]
碎碎念
这个教程, 应该相当的简单易懂了 ...吧?
明天就要过上两周才放假一次的生活了...... 啊啊啊啊啊啊......
