A Scalable Deep-Learning Framework for Generating and Grading Shape–Color Visual Reasoning Quizzes

Abstract :- A visual learning task is an assessment activity that uses images or other visual stimuli to engage and train learners’ perceptual and cognitive skills. By presenting information in a graphical form such as shapes, colors, diagrams, or spatial arrangements, it promotes the development of visual discrimination, pattern recognition, spatial reasoning, and memory by asking students to interpret, analyze, and respond to what they see. We address the lack of f lexible, large-scale, automatically gradable item banks by presenting an end-to-end framework that generates and evaluates multi-shape, multi-color visual reasoning quizzes using modern deep vision models. The pipeline constructs problems by arranging 1–5 distinct geometric shapes (circle, rectangle, triangle, pentagon, hexagon, star) in one of five colors (red, blue, green, yellow, black); each image is uniquely labeled, yielding a dataset of 2,000 items drawn from a 2.49 × 106 combinatorial space. Grading is posed as multi-label classification, and four pretrained backbones, ResNet-50, EfficientNet-B0, MobileNetV3-Large, and ViT-B/16 are f ine-tuned and evaluated with 5-fold cross-validation on accuracy, inference latency, and model size. Controlled synthesis provides clean ground truth and reproducibility, and a unified protocol enables fair comparisons. ViT-B/16 attains perfect mean accuracy, ResNet-50 reaches 99.9%, and lightweight CNNs exceed 95%, indicating that pretrained classifiers can reliably automate shape–color assessment and offer a scalable tool for visual reasoning education as well as a reproducible testbed for model evaluation.