Handwritten morse code recognition using convolutional neural networks

Even after many decades, Morse code has not disappeared from practical use. Still, recognising it in handwritten form is tricky, as each person’s writing style, spacing, and stroke patterns can differ greatly. The use of deep learning techniques for this purpose has been largely underexplored. In this paper, we evaluate five convolutional neural network (CNN) architectures—standard CNN, VGG16, ResNet50, MobileNet, and MobileNetV2—to address this challenge. These architectures were chosen to represent a range of model complexities and design philosophies—from lightweight mobile networks to deep residual and classical convolutional models. Three datasets were employed: (i) a manually curated dataset collected from 17 individuals with diverse handwriting styles and varying familiarity with Morse code, (ii) an external dataset sourced from the literature, and (iii) a combined dataset comprising approximately 38,976 images with extensive data augmentation. While the standard CNN was trained from scratch, transfer learning was applied to the four established models. ResNet50 consistently achieved the highest accuracies (95.9%,92.4%, and 93.0%) across the three datasets, while the standard CNN also produced competitive results (86.3%,91.6%, and 89.8%), outperforming the remaining models. These highly accurate models could be particularly valuable in educational contexts, where manually correcting learners’ guesses can be time-consuming.