Captcha Generation and Recognition Using a Convolutional Neural Network – Project Overview and Implementation

The project code is divided into three parts: Generate_Captcha (generates training, validation, and test captcha images and reads image data and labels), cnn_model (the convolutional neural network), and driver (model training and evaluation).

1. Configuration

<code>class Config(object):
    width = 160  # 验证码图片的宽
    height = 60  # 验证码图片的高
    char_num = 4  # 验证码字符个数
    characters = range(10)  # 数字[0,9]
    test_folder = 'test'    # 测试集文件夹，下同
    train_folder = 'train'
    validation_folder = 'validation'
    tensorboard_folder = 'tensorboard'  # tensorboard的log路径
    generate_num = (5000, 500, 500)  # 训练集，验证集和测试集数量
    alpha = 1e-3  # 学习率
    Epoch = 100  # 训练轮次
    batch_size = 64     # 批次数量
    keep_prob = 0.5     # dropout比例
    print_per_batch = 20    # 每多少次输出结果
    save_per_batch = 20     # 每多少次写入tensorboard</code>

2. Generate Captcha (class Generate)

Provides methods check_path() to ensure required folders exist and gen_captcha() to create captcha images, storing them in the appropriate directories.

Example captcha image:

3. Data Loading (class ReadData)

read_data() returns a numpy.array of images and their labels (file names). label2vec() converts a label string into a one‑hot vector. Example conversion:

<code>label = '1327'
label_vec = [0,1,0,0,0,0,0,0,0,0,
            0,0,0,1,0,0,0,0,0,0,
            0,0,1,0,0,0,0,0,0,0,
            0,0,0,0,0,0,0,1,0,0]</code>

load_data() loads all images from a folder and returns the image array, labels, and count.

4. Model Definition (cnn_model)

The network uses three convolutional layers, each with a filter size of 5. To mitigate over‑fitting, a dropout layer follows every convolution. The final layers reshape the feature maps into a matrix suitable for classification.

Model architecture illustration:

5. Training & Evaluation

next_batch() provides an iterator that yields data in batches. feed_data() feeds a batch to the model. The inputs are:

x : image array

y : image labels

keep_prob : dropout keep probability

evaluate() assesses the model on validation and test sets. run_model() orchestrates the full training‑and‑evaluation loop.

6. Current Results

After about 4,000 training iterations the model reaches >99% accuracy on the training set and ~93% on the test set, though slight over‑fitting is observed. Training runs on a CPU and takes roughly four hours per full run.

<code>Images for train ：10000, for validation : 1000, for test : 1000
Epoch : 1
Step     0, train_acc:   7.42%, train_loss:  1.43, val_acc:   9.85%, val_loss:  1.40, improved:*
Step    20, train_acc:  12.50%, train_loss:  0.46, val_acc:  10.35%, val_loss:  0.46, improved:*
... (omitted intermediate steps) ...
Epoch : 51
Step  7860, train_acc: 100.00%, train_loss:  0.01, val_acc:  92.37%, val_loss:  0.08, improved: 
Step  7880, train_acc:  99.61%, train_loss:  0.01, val_acc:  92.28%, val_loss:  0.08, improved: 
Step  7900, train_acc: 100.00%, train_loss:  0.01, val_acc:  92.42%, val_loss:  0.08, improved: 
Step  7920, train_acc: 100.00%, train_loss:  0.00, val_acc:  92.83%, val_loss:  0.08, improved: 
No improvement for over 1000 steps, auto-stopping....
Test accuracy:  93.00%, loss:  0.08</code>

7. TensorBoard

Before each training run the TensorBoard log directory is cleared to keep the visualizations tidy. Accuracy and loss curves are displayed as follows:

- END -