Gravitationally lensed quasars are important objects in astronomy for probing the universe. Unfortunately, these objects are exceedingly rare, occurring only for only ∼ 1/10 000 quasars. The challenge is to find these lensed quasars amongst large astronomical data sets. In contrast to previous attempts, which have only made use of numeric data, we perform semi-supervised classification based on images of quasars. These images are low resolution and noisy, but are enough for experienced astronomers to perform classification. Using virtual adversarial training to take advantage of millions of unlabelled images, we develop a classifier which achieves an F1 score of 0.49 — an extremely impressive result in this domain. Predictions made by this classifier are already being used to select candidates for telescopes around the world.