How Does EfficientRDet Work, the New Brain for Spotting Ships in Satellite Images?

Detecting ships in satellite images is trickier than it looks. Vessels can appear in any direction, against cluttered backgrounds, and at vastly different scales. To tackle this, researchers developed EfficientRDet, an algorithm built on the well‑known EfficientDet object detector.

EfficientRDet adds an angle‑prediction branch, letting the model understand whether a ship points at 30°, 120° or 270°. Then a “pseudo‑two‑stage” process kicks in: first a Rotated‑Bounding‑Box Refinement Branch (RRB) fine‑tunes the rotated box, followed by a Class‑Score Refinement Branch (CRB) that sharpens the confidence score.

Additional upgrades include an Enhanced BiFPN (E‑BiFPN) for faster feature fusion, an Attention Head that focuses on the most relevant image parts, and a Distribution Focal (DF) angle representation that reduces errors.

Tests on two datasets – HRSC2016 (optical) and RSDD‑SAR (radar) – deliver impressive numbers: 97.60% AP50 on HRSC2016 and 93.58% AP50 on RSDD‑SAR, beating the current state‑of‑the‑art. Ablation studies confirm each new component adds real value.

In short, EfficientRDet offers precise, fast ship detection, paving the way for more reliable and affordable maritime monitoring.

EfficientRDet builds upon EfficientDet D3, which uses an EfficientNet‑B3 backbone to extract multi‑scale features. The angle prediction branch employs a sinusoidal encoding of the angle, turning regression into a classification over 180 classes (1° per class). The Rotated‑Bounding‑Box Refinement Branch (RRB) uses a rotated IoU loss to optimize box overlap, while the Class‑Score Refinement Branch (CRB) applies a weighted focal loss to handle class imbalance between ships and background. The Enhanced BiFPN (E‑BiFPN) adds extra bidirectional connections and batch normalization to improve gradient flow. The Attention Head combines channel‑wise and spatial‑wise attention (CBAM) to highlight regions of interest. The DF angle representation uses a discrete Gaussian distribution, allowing the model to learn angular uncertainty and reduce over‑fitting. Ablation studies reveal that removing the angle branch drops AP50 by ~4%, while omitting E‑BiFPN costs ~2.5% precision. The network is trained with AdamW (lr = 1e‑4, weight decay = 0.01) for 120 epochs, using rotation‑ and scale‑specific data augmentation, ensuring robustness to orientation and size variations.