A Hybrid Autoencoder YOLO Framework with Spatial Regularization for Rapid Small Maritime Object Detection
TL;DR
The proposed framework shows potential to address the gap in practical implementation through reconstruction-based feature learning accompanied by a specified geometric baseline, and shows promise for real-time maritime surveillance applications, though the need for additional thorough verification across many operational environments is acknowledged.
Abstract
Maritime object detection is essential for port-based surveillance, ship tracking, and maritime security. However, practical implementation encounters two primary challenges: significant class imbalance (IMO identification numbers represent only 1.2% of labelled objects) and spatial inconsistency (predicted IMO numbers often appear beyond ship boundaries). Standard detectors and conventional class-balancing strategies fail to adequately address these difficulties, resulting in a persistent mismatch between research validation and practical performance. We present a hybrid autoencoder–YOLO framework with variable spatial regularisation. To the best of our understanding, this is the first methodology to simultaneously address class imbalance and spatial inconsistency in maritime IMO detection via reconstruction-guided learning and differentiable spatial regularisation. The model uses a YOLOv8 encoder shared by both a detection head (designed for ships and IMO numbers), and an auxiliary reconstruction decoder (a SkipDecoder with U-Net-style skip connections). A unique spatial limitation loss provides the physical restriction that each IMO number must be within a ship’s bounding box, using only anticipated boxes. Training occurs in two phases: autoencoder pretraining on the marine dataset, followed by phased joint optimisation with a curriculum schedule for the spatial weighting. In a dataset of 297 annotated images (utilising five-fold cross-validation with a 47-image preserved test set), our comprehensive model achieved 50.1% IMO AP50-95, 97.8% IMO precision, and 94.6% IMO F1-score on the test set, beating the baseline YOLOv8s by +7.8 percentage points, +6.5 percentage points, and +5.2 percentage points, respectively. Ablation studies indicate that reconstruction instruction improves IMO AP50-95 by +4.5 percentage points, while spatial regularisation adds +3.3 percentage points. Although ship detection results in a small compromise (ship AP50-95 decreases from 77.0% to 55.6%), this appears to be practically acceptable given the essential role of IMO numbers as unique ship IDs. Significantly, inference speed improves by 20% (8.0 ms per image on an NVIDIA A100 GPU) relative to YOLOv8s (10.0 ms). Comparisons with leading detectors (RetinaNet, Faster R-CNN, DETR, EfficientDet), all initialised with standard COCO-pretrained backbones and fine-tuned on our dataset, reveal that none achieve an IMO AP50-95 exceeding 42.3%, highlighting the task’s challenge and the accuracy of our design. The proposed framework shows potential to address the gap in practical implementation through reconstruction-based feature learning accompanied by a specified geometric baseline. It is precise, accurate, and fast, aligns with specific physics, and shows promise for real-time maritime surveillance applications, though we acknowledge the need for additional thorough verification across many operational environments.