Identifying Planetary Nebulae Candidates from IPHAS and VPHAS+ Surveys via YOLO Deep Learning Models

We present a deep learning detection method capable of effectively identifying planetary nebulae (PNe) on IPHAS and VPHAS+ surveys. Our detection pipeline is based on the YOLO architecture and is adaptable to other datasets. Using the HiPS system, we preprocess and augment the $r$, $i$, and $H{\alpha}$ bands to produce two red, green, and blue (RGB) image datasets: (1) $640 \times 640$ px images at full resolution, and (2) $1 280 \times 1 280$ px images designed to capture extended objects up to $0.2$ deg. Images of $\sim 2600$ PNe are labelled using the HASH database. The loss function and evaluation metrics are adapted to the diffuse and low-contrast nature of PNe; a very low intersection over union (IoU) threshold ($1\%$) is used. A $k$-fold cross validation shows that the YOLOv12m model achieves an average precision of $0.979$ on the first dataset. At a confidence threshold of $0.15$, it achieves a precision rate of $98.7\%$ and a recall rate of $97.4\%$. We deployed our system (on $1280 \times 1280$ images) on the full surveys' coverage ($\sim 3755$ deg$^2$), and present a preliminary analysis of the detection results. The system is robust against artifacts such as CCD corner defects and HiPS tile joints. Our convolutional-neural-network-based model achieved excellent results on the test set, comparable to-or exceeding-that of a recent Swin Transformer approach. However, a challenge persists when applying deep learning models to extensive and densely populated Galactic fields: the development of an automated system that can effectively discriminate PNe from mimics that human experts are able to distinguish (e.g., H II regions). We outline possible improvements for future work.