Planetary nebulae: a YOLO detector scans 3,755 square degrees of the galactic plane with 98.7 per cent accuracy

Trained on around 2,600 known nebulae from the IPHAS and VPHAS+ surveys, a model developed by a French team achieves a recall of 97.4 per cent, but still struggles with HII regions that resemble them.

Planetary nebulae, envelopes of gas ejected by stars at the end of their lives, are diffuse and low contrast, making them difficult to locate automatically. A French team has adapted the YOLO object detection architecture to the IPHAS and VPHAS+ galactic plane surveys, by combining the r, i and H-alpha bands into color images via the HiPS system.

Metrics adapted to blurred objects

Two sets of images were produced: 640 × 640 pixels at full resolution and 1,280 × 1,280 pixels to capture objects extended to 0.2 degrees. Approximately 2,600 nebulae from the HASH database were used as examples, and the overlap threshold between detection and label was lowered to 1% to account for their wispy contours. In cross-validation, YOLOv12m achieves an average accuracy of 0.979; at the confidence level of 0.15, it achieves 98.7% accuracy and 97.4% recall, at or above a recent Swin Transformer-based approach.

Deployed over the entire coverage of the surveys, i.e. about 3,755 square degrees, the system is resistant to artifacts such as sensor corner defects or tile joints. One obstacle remains in very dense galactic fields: distinguishing real nebulae from objects that resemble them, such as HII regions, which human experts are even better at.