This paper explores the application of transfer learning using the lightweight SqueezeNet architecture for origami image classification, with the focus on difficulty estimation. The study uses a dataset of 1,508 origami images labeled as easy, intermediate, or complex. The SqueezeNet model achieved a validation accuracy of 84.72%, with strong F1-scores of 91.08%, 75.00%, and 81.67% for the respective difficulty levels. When compared to VGG16-based benchmarks, SqueezeNet demonstrated superior efficiency and precision, particularly in handling complex geometric patterns typical in origami. Data augmentation and adaptive learning rate strategies further enhanced the model’s robustness and generalization. These strategies enhanced performance and also reduced overfitting. The findings of this study highlight SqueezeNet’s effectiveness as a resource-efficient tool for origami classification, as a basis for future advancements in machine learning applications within the domain of art and design. This work contributes to the field by offering a practical approach to classifying images with varying levels of complexity, showing the potential of deep learning in areas beyond traditional image recognition tasks.