A new research study has introduced YOLOv5-BSD, an optimized object detection algorithm specifically engineered to address the persistent challenges in maritime remote sensing imagery. Traditional target detection methods often struggle in marine environments due to complex, cluttered backgrounds, the diminutive scale of many ship targets, and dense target distributions where vessels are frequently clustered together. By modifying the standard YOLOv5 model architecture, the research team developed a specialized framework that integrates a Bidirectional Parallel Feature Pyramid Network (BiFPN), an enhanced loss function based on the SIoU metric, and deformable convolution modules. These advancements allow the model to more effectively capture fine-grained details while simultaneously managing the geometric variations typical of ship targets in satellite-derived images.
The methodology focused on three primary architectural enhancements to improve performance. First, the replacement of the traditional feature pyramid network with a BiFPN structure allowed for more efficient bidirectional cross-connection of features across different layers. This modification was crucial for digging deeper into shallow detail information without requiring a significant increase in the size of the original images. The researchers also added a dedicated detection layer specifically for small targets using the 4x downsampled feature map, which enabled the model to better distinguish small objects that were previously obscured by the deep-layer feature loss often occurring after extensive convolution and pooling processes.
The study also incorporated the SIoU loss function to refine the model’s target identification capabilities. Unlike traditional loss functions that primarily focus on overlapping area, distance, and aspect ratio, the SIoU metric reevaluates the vector angle between the predicted and ground truth bounding boxes. By introducing an angle-aware penalty, the researchers improved both training speed and inference accuracy, reducing the occurrence of random prediction boxes and improving the convergence process during training. This component effectively aligns the regression targets more precisely with the geometric reality of the observed ships, which often exhibit large aspect ratios.
To handle the geometric deformation caused by varied viewing angles, sizing, and ship modeling, the research team utilized version 2 of the deformable convolution algorithm. Standard convolution operators are limited to fixed sampling geometries, making them poorly suited for the high-aspect-ratio targets common in satellite remote sensing. The deformable convolution module, however, allows the network to learn horizontal and vertical offsets during the sampling process. By learning these new weights, the model can adapt its receptive field to the actual shape of the target, successfully overcoming the missed detections and false positives that plague models limited by rigid, square-shaped kernels.
Experimental evaluation of the YOLOv5-BSD algorithm was performed on the ShipRSImageNet V1.0 data set, which comprises 3,435 images and 17,573 ship instances across 50 categories. The dataset was divided into 2,198 training images, 550 validation images, and 687 test images. Conducted on an RTX 4090 GPU using the PyTorch framework, the training process operated with a batch size of 64 over 300 epochs. Performance was measured using standard metrics including precision, recall, and mean average precision (mAP) at various intersection over union (IoU) thresholds, providing a comprehensive assessment of the model’s reliability in identifying maritime vessels against complex backdrops.
The results showed that YOLOv5-BSD demonstrated superior performance compared to the original YOLOv5s baseline. Specifically, the improved model achieved an increase of 2.5 percentage points in mAP@0.5 and 3.7 percentage points in mAP@0.5:0.95. Furthermore, precision increased by 2.48 percentage points and recall improved by 2.15 percentage points. These metrics verify the effectiveness of the integration strategy, demonstrating that the combination of BiFPN, the optimized loss function, and deformable convolution significantly enhances the sensitivity of the model to both dense clusters of small ships and complex environments where buildings or shore structures might otherwise interfere with detection accuracy.
A detailed ablation analysis confirmed that each module contributed positively to the overall performance gains. Introducing the BiFPN and small object detection layer improved mAP@0.5 by 1.2 percentage points alone, while the SIoU loss function added another 0.4 percentage points. The incorporation of deformable convolution provided the most notable jump in performance, increasing precision by 1.2 percentage points and recall by 0.96 percentage points. Importantly, these advancements also resulted in a more computationally efficient detection process, achieving high performance metrics with only 16.7% of the computational requirements (GFLOPs) of the large YOLOv5x model.
While the YOLOv5-BSD model outperformed the original baseline and showed competitive results against newer, larger models like YOLOv8s and YOLO-X, the researchers noted remaining limitations. Analysis of the confusion matrix revealed that certain categories, such as sailboats, are still frequently ignored or misclassified. The study also observed that even when target localization is accurate, some vessels are mislabeled as general “other ship” categories. These findings highlight the inherent challenges of fine-grained classification within dense, multi-category maritime datasets where inter-class differences are minimal.
In summary, the YOLOv5-BSD algorithm provides a robust framework for maritime remote sensing, demonstrating that targeted optimizations—rather than simply relying on massive parameter scaling—can yield high-performance target detection. By enhancing feature fusion, refining loss calculation, and introducing deformable sampling, the researchers successfully improved the detection of ships against cluttered backgrounds and within high-density clusters. Future research directions will focus on addressing the current classification weaknesses, specifically improving the recognition accuracy for sailboats and reducing intra-class classification errors within diverse maritime environments.
The research utilized the ShipRSImageNet V1.0 dataset, a repository specifically curated for maritime satellite remote sensing, to benchmark the new algorithm. This dataset comprises more than 3,435 images containing over 17,573 ship instances across 50 distinct categories. The expert-led annotation process utilized both horizontal and oriented bounding boxes to capture the ships in various orientations, providing a robust ground truth for training deep learning models. The data was split into 2,198 training, 550 validation, and 687 testing images to ensure comprehensive evaluation. Crucially, the dataset reflects the practical realities of remote sensing, including varying weather conditions, diverse ship models, and differing locations. Statistical analysis reveals that a significant proportion of these vessels are small, occupying only a few dozen pixels within the high-resolution imagery. Furthermore, the dataset exhibits class imbalance, with certain categories like “other ship” and “other warship” significantly more represented than others. This high-density distribution and the challenge of fine-grained classification where inter-class differences are minimal provide a rigorous testing ground for detection algorithms, distinguishing the performance of specialized architectures from more generic, off-the-shelf detection frameworks.
The structural overhaul centered on replacing the conventional feature pyramid network—which utilized a cascaded architecture of FPN and PANet—with a Bidirectional Parallel Feature Pyramid Network (BiFPN). In the baseline YOLOv5s model, while the FPN-PANet architecture successfully facilitates feature transmission and fusion, the inherent downsampling processes and expanding receptive fields often cause overlapping, leading to overly fine-grained feature maps. This granularity inadvertently discards essential details found in shallow feature maps, which are critical for the precise localization of small objects. By contrast, the BiFPN structure enables interweaving features in both top-down and bottom-up directions, creating a more robust fusion pathway. The implementation specifically removed nodes with only a single input edge to streamline the network without sacrificing performance. Additionally, the developers introduced an extra edge between original input and output nodes in the same layer to facilitate broader feature integration without adding prohibitive computational costs. This architecture also integrates the 4x downsampled feature map, which the original model bypassed. By focusing on these shallow detail maps and using a dedicated P2 branch, the model can retain critical semantic information that might otherwise be lost through excessive convolution and pooling, significantly enhancing the detection sensitivity for clustered small vessels.
A core refinement to the training process involved substituting standard box loss functions with the SIoU metric. Previous loss functions were limited in their focus, typically assessing only the distance, overlapping area, and aspect ratio between the predicted and ground-truth bounding boxes without accounting for the directional relationship. The SIoU loss function introduces a reevaluation of the vector angle between expected regressions, which is computed using a four-part penalty metric: angle cost, distance cost, shape cost, and IoU loss cost. The introduction of the angle-aware loss function is designed to minimize the variables involved in distance-related uncertainty. During the training convergence process, the model prioritizes minimizing the vector angle to align the predicted regression box with the ground truth along the X or Y axis before minimizing distance. This angle cost, defined through trigonometric functions, effectively reorients the model’s prediction process. The researchers found that by setting the theta parameter to four—within a suggested range of two to six—they could optimize shape without compromising the flexibility of the regression. This refinement not only accelerates training speed but also improves inference accuracy by drastically reducing the occurrence of anomalous, randomly placed prediction boxes that often characterize standard detection failures in dense remote sensing datasets.
To further address the geometric challenges inherent in maritime imagery, the team implemented deformable convolution v2. Standard convolutional neural networks rely on a fixed sampling grid, where the receptive field for every point remains uniform. This rigidity is fundamentally ill-suited for ship targets, which frequently exhibit large aspect ratios and variable rotations in satellite observations. Deformable convolution v2 allows for a dynamic receptive field by learning horizontal and vertical offsets during the sampling process, which are applied to the standard grid. The “v2” enhancement specifically addresses potential interference from regions unrelated to the target by learning an additional weight parameter. This allows the convolution operator to restrict its coverage to relevant pixels, effectively suppressing noise from the surrounding environment. By adjusting the sampling positions to better match the elongated shape of a ship, the network can accommodate geometric deformations resulting from camera viewing angles, size differences, and target modeling. This integration proved essential for correcting the common problem where rigid kernels failed to capture the full extent of an object, leading to either missed targets or false detections. The experimental results confirmed that this module not only improved detection accuracy—with precision, recall, and mAP metrics showing clear gains—but also contributed to greater overall computational efficiency compared to static, non-deformable kernels. This modular approach ensures that the model remains robust against the natural visual variability encountered in aerial maritime surveillance.
Subject of Research: Optimization of YOLOv5 for improved ship detection in remote sensing maritime imagery.
Article Title: Yolov5-BSD: a maritime remote sensing object detector based on Yolov5
Article References: Liu, T., Jiang, X., Cen, Q., & Song, T. (2026). Yolov5-BSD: a maritime remote sensing object detector based on Yolov5. Journal on Image and Video Processing, 2026(1), Article 7. https://doi.org/10.1186/s13640-026-00690-x
Image Credits: AI Generated
DOI: 10.1186/s13640-026-00690-x
Keywords: Object detection, Remote sensing, Maritime surveillance, YOLOv5, Computer vision, Deformable convolution, Feature fusion, Satellite imagery, BiFPN, SIoU loss function, Deep learning, Marine environment, Small target detection, Image processing, ShipRSImageNet, Algorithm optimization, Maritime safety, Target classification
Cite Scienmag News
Blake Davidson. (November 18, 2026). Optimized AI Model Significantly Enhances Maritime Remote Sensing Object Detection. Scienmag. https://scienmag.com/optimized-ai-model-significantly-enhances-maritime-remote-sensing-object-detection/
Blake Davidson. "Optimized AI Model Significantly Enhances Maritime Remote Sensing Object Detection." Scienmag, 18 November 2026, https://scienmag.com/optimized-ai-model-significantly-enhances-maritime-remote-sensing-object-detection/. Accessed 18 November 2026.
Blake Davidson. "Optimized AI Model Significantly Enhances Maritime Remote Sensing Object Detection." Scienmag. November 18, 2026. https://scienmag.com/optimized-ai-model-significantly-enhances-maritime-remote-sensing-object-detection/








