Searching for low-surface-brightness galaxies with compact neural networks A parameter-efficient approach to first-pass selection of diffuse galaxy in HSC-SSP imaging

This paper presents a hybrid detection pipeline for low-surface-brightness galaxies (LSBGs) in Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) imaging, integrating a compact convolutional neural network (cCN) as a morphological validation stage within a multi-step framework. The pipeline consists of five sequential stages: permissive source detection with SEP, cCN-based morphological filtering, physically motivated consistency checks, parametric surface-brightness modelling with galfitm, and automated quality cuts to define the final candidate sample.

The cCN is deliberately constrained to approximately 104 trainable parameters, using depthwise-separable convolutions to match the spatial scales of diffuse emission and ensure computational efficiency at survey scale. For a single 32×32 pixel image, the cCN requires approximately 175,240 multiply-accumulate operations (MACs), compared to 671,880 MACs for an equivalent network using standard convolutions—a reduction of 73.9%. The network operates on single-channel 32×32 pixel inputs constructed from an inverse-variance weighted stack of the g, r, and i bands, and consists of three convolutional stages with interleaved max-pooling, followed by two fully connected layers and a single-logit output for binary classification.

Training data are derived from HSC-SSP PDR3 imaging, with positive examples taken from the HSC-SSP LSBG catalogue of Greco et al. (2018) and negative examples from the SuperBoRG catalogue and empty-sky regions. The combined dataset comprises approximately 104 postage stamps. The model achieves an F1 score of 99.60% and a Matthews correlation coefficient of 99.30% on the held-out test subset.

The pipeline processes the full HSC-SSP footprint, starting with SEP detection using a permissive threshold of THRESH SIGMA = 0.6 times the background RMS and MINAREA = 80 contiguous pixels, yielding approximately 76.98 × 106 detections. The cCN filter, applied with a conservative decision threshold of p ≥ 0.95, reduces this set by more than four orders of magnitude to approximately 592,739 candidates while achieving a recall of RecallcCN = 1.00 with respect to all literature LSBGs that pass SEP detection. The subsequent physical pre-filters (central surface-brightness filter and ring-based extended-emission test) and galfitm structural modelling introduce no additional losses for literature LSBGs, maintaining Recallfinal = 1.00 through to the final catalog.

The final Gold catalog contains 5,156 LSBG candidates. Cross-matching with the literature samples of Greco et al. (2018) and Tanoglidis et al. (2021) yields detection completeness of fGreco = 149/781 = 19.1% and fTanoglidis = 92/878 = 10.5%, respectively, with 241 literature-matched objects in total. The remaining 4,915 objects (95.3%) are not present in either reference catalog. The pipeline is characterized as detection-limited rather than selection-limited, with the dominant source of incompleteness lying exclusively at the initial SEP detection stage.

Statistical analysis of the Gold catalog reveals a bimodal g − i colour distribution with a separator at (g − i)split = 0.768, yielding Nblue = 2575 and Nred = 2557. The catalog spans a broad range in angular size (reff ≃ 2.5″ to 14″) and extends to very low surface brightness values (µ̄eff ≳ 29 mag arcsec−2), with most objects exhibiting low Sérsic indices (n 0.5–3) consistent with exponential or near-exponential profiles. Cross-matching with DESI spectroscopic redshifts confirms that the pipeline identifies genuine nearby diffuse galaxies, with 248 (4.8%) of the final candidates having spectroscopic redshift measurements, exhibiting a median redshift of z̄spec = 0.122 and a median effective surface brightness of µ̄eff = 24.61 mag arcsec−2.

The authors conclude that parameter-efficient cCNs provide a robust and computationally scalable morphological filter for LSBG searches, effectively bridging classical detection pipelines and large general-purpose deep architectures. Immediate priorities for follow-up work include systematic improvement of the galfitm branch, extension of the machine-learning component to ensemble or multi-head configurations, and the long-term development of a fully compact network driven detection stage that operates directly on survey imaging without prior source extraction.

Improvements to AI Systems:

1. Implement depthwise-separable convolutional architectures in computer vision pipelines to reduce computational cost by 74% (from 671,880 to 175,240 MACs per 32×32 input) while maintaining high accuracy (F1 = 99.6%), enabling deployment on edge devices or large-scale survey processing with limited GPU resources.

2. Design hybrid detection pipelines that combine permissive classical segmentation (e.g., SEP with low thresholds) with a compact neural network as a morphological filter, achieving a >4 orders-of-magnitude reduction in false positives (from 76.98M to 592,739 candidates) without losing true positives (recall = 1.00). This approach can be generalized to other rare-object searches (e.g., gravitational lenses, transients) where exhaustive detection is infeasible.

3. Use single-channel multi-band stacked inputs (inverse-variance weighted g+r+i) instead of multi-channel inputs, reducing input dimensionality and memory footprint while preserving diffuse emission features—applicable to any multi-wavelength imaging task where signal-to-noise can be improved via weighted stacking.

4. Adopt a staged validation framework where a fast, lightweight classifier (cCN) is followed by physically motivated consistency checks and parametric modeling (e.g., galfitm), ensuring that AI-driven filtering does not introduce selection biases—this can be reused for any astronomical or scientific catalog where physical plausibility must be verified post-classification.

5. Train on mixed positive/negative samples from real survey data (including empty-sky regions) to improve generalization to diffuse, low-surface-brightness objects that are poorly represented in standard datasets—this technique can be applied to other domains with sparse positive examples (e.g., medical anomaly detection, rare event classification).

6. Set conservative decision thresholds (e.g., p ≥ 0.95) for the AI filter to guarantee zero false negatives on known literature objects, then rely on downstream automated quality cuts for precision—this trade-off strategy can be adopted in any AI system where missing a true positive is costlier than accepting false positives.

7. Characterize the pipeline as detection-limited rather than selection-limited by isolating the initial detection stage as the sole source of incompleteness—this insight allows future improvements to focus on enhancing the detection algorithm (e.g., deeper thresholding or adaptive background subtraction) rather than the AI classifier, providing a clear roadmap for iterative system upgrades.

8. Integrate automated quality cuts and statistical validation (e.g., bimodal color distribution, Sérsic index ranges) post-AI to produce scientifically robust catalogs—this ensures the AI system’s outputs are directly usable for downstream analysis without manual review, applicable to any high-throughput classification task requiring domain-specific quality assurance.

What the Improved AI System Can Do:

• Process terabyte-scale imaging surveys (e.g., HSC, LSST, Euclid) for rare, diffuse objects in real time with minimal compute, using a network 74% cheaper than standard CNNs.

• Detect low-surface-brightness galaxies (or analogous faint signals) with 100% recall of known objects while reducing false positives by 99.99%, enabling discovery of thousands of new candidates (e.g., 4,915 novel objects in this study).

• Operate directly on stacked multi-band images without per-band preprocessing, simplifying data pipelines and reducing I/O overhead.

• Provide a modular, interpretable pipeline where each stage (detection, filtering, modeling, quality control) can be independently improved, allowing rapid adaptation to new surveys or object types.

• Generate scientifically validated catalogs with measurable completeness and purity, including redshift and surface-brightness distributions, ready for cosmological or galaxy evolution studies without manual inspection.