Adaptive optics scanning laser ophthalmoscope (AOSLO) imaging enables the cone photoreceptor mosaic to be visualised in the living human eye. Performing quantitative analysis of these images requires identification of individual photoreceptors. This is typically performed by manual labelling, which is subjective, time consuming and not feasible on a large scale. Automated algorithms to replace manual labelling are required and deep learning-based methods provide an effective way of achieving this. However, this approach requires large volumes of annotated training data that are difficult to acquire. Synthetic data may help to bridge this lack of annotated training data. A U-Net configuration was trained using a large synthetic dataset of confocal AOSLO images generated using ERICA alongside a smaller dataset of real confocal AOSLO images (Milwaukee dataset). Model performance was assessed by calculating the Dice coefficient, a metric quantifying segmentation overlap, on both a real held-out test set and an independent real dataset (Oxford dataset). Results from this evaluation were benchmarked against expert labelling and two automated cone detection methods: a confocal convolutional neural network (CNN) (1), and a combined graph-theory and dynamic programming approach (2)). The mean Dice coefficient compared to manual labelling was 0.989 (U-Net), 0.989 (confocal CNN), and 0.985 (graph-theory and dynamic programming) on the held-out test set. On the independent Oxford dataset, the U-Net achieved a mean Dice coefficient of 0.962 compared to manual labelling. Results show performance that is comparable to the gold standard of manual labelling and two automated cone detection methods. Furthermore, we demonstrate generalisability of this approach on an independent real dataset with images from higher retinal eccentricities. This approach may be useful for quantitative analysis of the photoreceptor mosaic in patients with retinal disease to provide cell-specific imaging biomarkers from AOSLO images.