An Efficient Multi-Character CAPTCHA Recognition System Using Convolutional Neural Networks

Abstract

With the increasing reliance on web-based services, ensuring secure authentication
mechanisms has become essential. CAPTCHA (Completely Automated Public Turing
test to tell Computers and Humans Apart) is widely used to prevent automated bots from
accessing online systems. However, advancements in machine learning and deep learning
have made it possible to break traditional CAPTCHA systems. This research presents an
efficient CAPTCHA recognition system using Convolutional Neural Networks (CNN)
capable of identifying multi-character CAPTCHA images with high accuracy.The
proposed system focuses on recognizing CAPTCHA images consisting of five
alphanumeric characters. Unlike traditional optical character recognition (OCR)
techniques, which rely heavily on segmentation and handcrafted features, this system
leverages deep learning to automatically extract features and classify characters. The
model processes grayscale images of size 50×200 pixels and predicts each character
using a multi-output CNN architecture.The architecture includes multiple convolutional
layers for feature extraction, followed by pooling layers to reduce spatial dimensions.
Batch normalization is applied to improve training stability and convergence. The
extracted features are flattened and passed through fully connected layers with dropout to
prevent overfitting. The model generates five separate outputs corresponding to each
character in the CAPTCHA, allowing simultaneous prediction of all characters.The
dataset is preprocessed by normalizing pixel values and encoding labels into categorical
format. The model is trained using categorical cross-entropy loss and optimized using the
Adam optimizer. Training performance is evaluated using accuracy and loss metrics, and
results are visualized using graphical plots.