The objective of this study is to compare convolutional neural networks and multilayer perceptron algorithms for enhancing image-based CAPTCHA systems to improve human-bot differentiation and CAPTCHA Security. Materials and Methods: The study uses a dataset of 32,000 images, categorized into two labels: Label 1 for human-drawn images and Label 2 for machine-generated images. The dataset underwent thorough cleaning using image preprocessing techniques and was split into training and testing subsets. Both convolutional neural networks (CNN) and multilayer perceptron (MLP) models were trained for 10 epochs with the training data further split into 20 batches. The statistical power is set at 80%, indicating an 80% chance of detecting an effect if it exists, and the confidence interval (CI) is set at 95%, reflecting a 95% level of certainty in the results. Results: The convolutional neural networks (CNN) model achieved an accuracy of 94.21%, significantly outperforming the multilayer perceptron (MLP) model, which achieved an accuracy of 56.42%.