Emotion recognition in children’s drawings provides valuable insights into their psychological and emotional states. This study compares two computational approaches for classifying positive and negative emotions in sketches drawn by children aged four to eight. The first approach employs classical machine learning techniques, where shape and intensity features are extracted from grayscale and binary images. These feature vectors are then evaluated using Support Vector Machines (SVM), Multilayer Perceptron (MLP), and Random Forest classifiers, with the best accuracy reaching 80% using SVM with a linear kernel. The second approach applies a pre-trained Vision Transformer (ViT) model, exploring the effects of data augmentation and fine-tuning on the classification performance. The ViT model achieved 76.5% accuracy with transfer learning. The results highlight the effectiveness of SVM and ViTs for analyzing children’s drawings, demonstrating that classical methods can perform competitively in small datasets. Meanwhile, deep learning models show promise but require more extensive datasets for optimal performance.