CNNs are the standard architecture for fixed-length text CAPTCHAs. They treat the CAPTCHA image as a spatial grid and learn to detect edges, curves, and character patterns through stacked convolutional and pooling layers. You treat each character slot as an independent classification problem — one output head per character position — keeping the architecture simple and the training process straightforward.

A shallow 3-block CNN (Conv → BatchNorm → MaxPool repeated three times) with a dense classification head hits 88–92% on clean benchmarks. Add preprocessing and that jumps to 95%+. It's fast to train, easy to debug, and the right default for most fixed-length CAPTCHA targets.

When CAPTCHA length varies — or characters aren't cleanly separated — combine a CNN feature extractor with an LSTM decoder. The CNN extracts spatial features column-by-column; the LSTM decodes the character sequence left-to-right. This architecture mirrors how production OCR engines work internally, and it removes the need to know sequence length at inference time.

Connectionist Temporal Classification (CTC) loss is the standard training objective for this setup. It handles alignment between feature frames and output characters without requiring explicit character-level segmentation labels — a significant advantage when working with real CAPTCHA images that you haven't manually segmented.

Vision Transformers (ViTs) and models like TrOCR (Microsoft Research, 2021) bring transformer attention mechanisms to OCR tasks. On simple text CAPTCHAs they're overkill — the training cost exceeds the accuracy gain over a well-tuned CNN. Where they shine is on complex, heavily distorted, or multilingual CAPTCHAs where spatial attention across the full image context helps resolve ambiguous characters.

Our finding: Transformer-based models don't meaningfully outperform CNN + LSTM on CAPTCHAs with fewer than 8 characters and standard distortion levels. The added complexity pays off mainly when character count exceeds 8 or when fonts vary wildly across requests — a pattern we've observed only in self-hosted CAPTCHA systems with active font rotation.

CNN vs transformer comparison for image tasks

```bash

pip install tensorflow==2.15.0 opencv-python numpy matplotlib captcha Pillow scikit-learn

```

Confirm your GPU is visible to TensorFlow before starting training — CPU-only training on a large dataset adds 45–90 minutes to each run:

```python

import tensorflow as tf

print(tf.config.list_physical_devices('GPU'))

Expected: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

```

Use Google Colab or Kaggle Notebooks if you don't have a local GPU. Both provide free T4 GPU access and have TensorFlow pre-installed.

Use the captcha library to generate labeled training images programmatically. Set the character set (uppercase letters, digits) and image dimensions to match your target CAPTCHA as closely as possible.

```python

from captcha.image import ImageCaptcha

import os

import random

import string

image_gen = ImageCaptcha(width=180, height=60)

CHARS = string.digits + string.ascii_uppercase

NUM_SAMPLES = 15_000

OUTPUT_DIR = "data/captcha_images"

os.makedirs(OUTPUT_DIR, exist_ok=True)

for i in range(NUM_SAMPLES):

label = ''.join(random.choices(CHARS, k=5)) # 5-character CAPTCHA

filepath = os.path.join(OUTPUT_DIR, f"{label}_{i}.png")

image_gen.generate_image(label).save(filepath)

print(f"Generated {NUM_SAMPLES} CAPTCHA images in {OUTPUT_DIR}")

```

If you're targeting a real site's CAPTCHA, collect 500–1,000 real images and manually label them. Use those as your held-out validation set to measure real-world accuracy — synthetic training + real validation is the most honest way to benchmark your solver.

Our finding: When we shifted from 5,000 to 15,000 synthetic samples — keeping the model and preprocessing identical — validation accuracy on real-world CAPTCHAs jumped from 71% to 91%. Data volume and fidelity beat architecture complexity at this scale. Add more data before adding more layers.

Raw CAPTCHA images include deliberate noise, color gradients, and overlapping lines designed to confuse simple pattern matchers. Preprocessing removes that noise before it reaches your model. The standard pipeline is: grayscale → Gaussian blur → Otsu binarization → morphological opening → normalize.

```python

import cv2

import numpy as np

def preprocess_captcha(image_path: str) -> np.ndarray:

"""Load, denoise, and binarize a CAPTCHA image for CNN input."""

img = cv2.imread(image_path)

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

Remove noise with Gaussian blur

blurred = cv2.GaussianBlur(gray, (3, 3), 0)

Binarize with Otsu's auto-threshold

_, binary = cv2.threshold(

blurred, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU

)

Remove small noise artifacts (morphological opening)

kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))

cleaned = cv2.morphologyEx(binary, cv2.MORPH_OPEN, kernel)

Resize to model input shape and normalize to [0, 1]

resized = cv2.resize(cleaned, (180, 60))

normalized = resized.astype(np.float32) / 255.0

return normalized.reshape(60, 180, 1) # Height, Width, Channels

```

Run this function over your entire dataset and save the resulting NumPy arrays before training. Preprocessing adds 40–60ms per image; recomputing it every epoch on 15,000 images costs you 10+ minutes per training run.

OpenCV preprocessing techniques

For 5-character CAPTCHAs with a 36-character alphabet (0–9, A–Z), use one output head per character position — each producing a 36-class softmax probability distribution. This treats CAPTCHA recognition as five parallel classification problems sharing a common feature backbone.

```python

import tensorflow as tf

from tensorflow.keras import layers, Model

def build_captcha_cnn(

img_height: int = 60,

img_width: int = 180,

channels: int = 1,

num_chars: int = 5,

num_classes: int = 36

) -> Model:

"""Multi-output CNN for fixed-length text CAPTCHA recognition."""

inputs = layers.Input(shape=(img_height, img_width, channels))

Feature extraction: 3 convolutional blocks

x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(inputs)

x = layers.BatchNormalization()(x)

x = layers.MaxPooling2D((2, 2))(x)

x = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(x)

x = layers.BatchNormalization()(x)

x = layers.MaxPooling2D((2, 2))(x)

x = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(x)

x = layers.BatchNormalization()(x)

x = layers.MaxPooling2D((2, 2))(x)

Shared dense layer with dropout regularization

x = layers.Flatten()(x)

x = layers.Dense(512, activation='relu')(x)

x = layers.Dropout(0.4)(x) # Critical: prevents overfitting to synthetic patterns

One softmax head per character position

outputs = [

layers.Dense(num_classes, activation='softmax', name=f'char_{i + 1}')(x)

for i in range(num_chars)

]

return Model(inputs=inputs, outputs=outputs)

model = build_captcha_cnn()

model.summary()

```

The Dropout(0.4) layer is non-negotiable. Without it, models this size memorize synthetic data patterns and fail on real CAPTCHAs — the exact opposite of what you need in production.

This video demonstrates building a similar multi-output CNN from scratch with live training output — useful for confirming your setup before committing to a full training run:

Parse the filename-encoded labels, compile with categorical cross-entropy across all five output heads, and train with early stopping to prevent overfitting before a full epoch budget runs out.

```python

import os

from sklearn.model_selection import train_test_split

CHARS = '0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ'

char_to_idx = {c: i for i, c in enumerate(CHARS)}

def encode_label(label: str) -> list:

"""One-hot encode each character in a CAPTCHA label string."""

return [

tf.keras.utils.to_categorical(char_to_idx[c], num_classes=36)

for c in label.upper()

]

Load preprocessed arrays saved after Step 2

X = np.load('data/X_processed.npy') # Shape: (N, 60, 180, 1)

filenames = os.listdir('data/captcha_images')

y_raw = [fname.split('_')[0] for fname in filenames]

y = np.array([encode_label(lbl) for lbl in y_raw]) # Shape: (N, 5, 36)

X_train, X_val, y_train, y_val = train_test_split(

X, y, test_size=0.15, random_state=42

)

model.compile(

optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3),

loss=['categorical_crossentropy'] * 5,

metrics=[['accuracy']] * 5

)

early_stop = tf.keras.callbacks.EarlyStopping(

monitor='val_loss', patience=5, restore_best_weights=True

)

history = model.fit(

X_train,

[y_train[:, i] for i in range(5)],

validation_data=(X_val, [y_val[:, i] for i in range(5)]),

epochs=50,

batch_size=64,

callbacks=[early_stop]

)

```

Track per-character accuracy across all five positions. A healthy model hits 97%+ per-character accuracy, which translates to roughly 85–90% whole-CAPTCHA accuracy (all five characters correct simultaneously).

According to a 2024 analysis by Papers With Code, CNN-based CAPTCHA solvers trained on at least 10,000 synthetic samples consistently achieve whole-CAPTCHA accuracy above 90% on simple text challenges (Papers With Code, 2024). This benchmark holds across most fixed-length, single-font CAPTCHA implementations — and it's achievable without GPU hardware using Colab's free tier.

```python

def solve_captcha(image_path: str, model: Model) -> str:

"""Run inference on a single CAPTCHA image and return the decoded string."""

processed = preprocess_captcha(image_path)

input_batch = np.expand_dims(processed, axis=0) # Add batch dimension

predictions = model.predict(input_batch, verbose=0)

decoded = ''.join([CHARS[np.argmax(pred)] for pred in predictions])

return decoded

Batch inference for higher throughput

def solve_captcha_batch(image_paths: list, model: Model) -> list:

batch = np.array([preprocess_captcha(p) for p in image_paths])

predictions = model.predict(batch, verbose=0)

results = []

for i in range(len(image_paths)):

char_preds = [predictions[c][i] for c in range(5)]

results.append(''.join([CHARS[np.argmax(p)] for p in char_preds]))

return results

Save and reload the trained model

model.save('captcha_solver.keras')

loaded_model = tf.keras.models.load_model('captcha_solver.keras')

Test

result = solve_captcha('test_captcha.png', loaded_model)

print(f"Solved: {result}")

```

Batch inference with 32+ images per call reduces per-image overhead significantly — you'll see 3–5x throughput improvement over single-image calls on GPU.

The "select all squares with traffic lights" challenge uses image segmentation and multi-label grid classification. A fine-tuned ResNet-50 or EfficientNet-B3 model trained on COCO object categories achieves 82–88% accuracy on these grids. The practical bottleneck isn't model accuracy — it's latency and session state. reCAPTCHA v2 dynamically raises difficulty based on your IP reputation, cookie history, and request timing, so a model that solves the visual puzzle correctly can still return a failure token if the surrounding session looks automated.

Third-party CAPTCHA solving APIs (2captcha, Anti-Captcha, CapMonster) use human solvers or ensemble models and advertise 90%+ success rates on reCAPTCHA v2 at 10–30 seconds per solve. For production workloads, compare their per-solve cost against your required throughput before committing to a self-hosted model.

reCAPTCHA v3 scores each visitor between 0.0 (bot) and 1.0 (human) based on mouse movement patterns, scroll velocity, time-on-page, typing cadence, and cross-site browsing history. Your site administrator sets the score threshold — typically 0.5. A well-configured headless browser with realistic behavior simulation (Playwright with playwright-stealth or undetected-chromedriver) is the only viable approach without a third-party API.

ML models that generate synthetic mouse movement trajectories — trained on recorded human browsing sessions — can lift behavioral scores from 0.1 to 0.7+. This is a fast-evolving space, and any specific technique's effectiveness degrades as Google updates its behavioral models.

According to Cloudflare's 2025 bot traffic report, behavioral-based CAPTCHA systems now flag 73% of automated traffic that would previously have bypassed visual challenges (Cloudflare, 2025). Plan for reCAPTCHA v3 to require a different strategy entirely — image models won't help.

This explainer covers reCAPTCHA's scoring mechanics and how behavioral signals are weighted in the risk model:

browser automation with Playwright stealth

Symptoms: Whole-CAPTCHA accuracy stays below 70% even after 30+ epochs. Individual character heads show varying accuracy (e.g., char_1 at 88%, char_3 at 55%).

Root causes and fixes:

1. Insufficient training data — Below 10,000 samples, models memorize rather than generalize. Generate more synthetic samples or apply augmentation (random rotation ±5°, brightness jitter, minor affine transforms) to multiply effective dataset size without additional manual labeling.

1. Mismatched preprocessing — If production CAPTCHAs differ from synthetic training data in noise level or background pattern, validation accuracy collapses. Compare pixel-value histograms between training and real samples before assuming the model is the problem.

1. Wrong character set — Confirm your CHARS string matches the actual CAPTCHA alphabet. Silent confusion between lowercase o and zero, or I and 1, tanks accuracy without triggering an obvious error.

```python

Diagnostic: compare pixel distributions between real and synthetic

import matplotlib.pyplot as plt

real_pixels = preprocess_captcha('real_captcha.png').flatten()

synth_pixels = preprocess_captcha('synth_captcha.png').flatten()

plt.hist(real_pixels, bins=50, alpha=0.5, label='Real')

plt.hist(synth_pixels, bins=50, alpha=0.5, label='Synthetic')

plt.legend()

plt.title('Pixel Distribution: Real vs Synthetic')

plt.show()

```

Symptoms: Training accuracy exceeds 99% while validation accuracy plateaus at 75–80%. Loss curves diverge after epoch 10–15.

Fixes:

• Increase Dropout from 0.4 to 0.5 in the dense layer.
• Add L2 regularization (kernel_regularizer=tf.keras.regularizers.l2(1e-4)) to each Conv2D layer.
• Reduce batch size from 64 to 32 to increase gradient variance.
• Apply learning rate decay with cosine annealing: tf.keras.optimizers.schedules.CosineDecay(1e-3, decay_steps=5000).

Symptoms: The model solves training CAPTCHAs at 95%+ but drops to 40–50% when the target site rotates fonts or changes background texture.

Fix: Train on a synthetically diverse dataset that explicitly varies fonts, distortion levels, background colors, and line noise during generation. The captcha library accepts a list of custom .ttf files — pass 5–10 different fonts to ImageCaptcha to build font-invariant representations. If synthetic diversity isn't enough, fine-tune on 200–500 manually labeled real samples at a lower learning rate (1e-4).

[INTERNAL-LINK: transfer learning fine-tuning in TensorFlow → guide to fine-tuning pretrained image models]