LSFDNet: Fusion for Maritime Ship Detection

• The paper introduces LSFDNet as a single-stage, end-to-end network that jointly optimizes image fusion and object detection for maritime ship detection.
• LSFDNet leverages complementary SWIR and LWIR imaging to enhance detection in low illumination, haze, fog, and complex nearshore backgrounds.
• Empirical results on the NSLSR dataset show LSFDNet achieving superior precision, recall, and mAP, with notable gains in strict localization metrics.

Searching arXiv for LSFDNet and closely related naming variants to ground the article in the relevant papers. LSFDNet denotes the Long-Short Wave Fusion Detection Network, a single-stage fusion and detection network for maritime ship detection from registered short-wave infrared (SWIR) and long-wave infrared (LWIR) image pairs. It was introduced for complex nearshore conditions involving low illumination or nighttime, haze, fog, smoke, sea-surface clutter, complex backgrounds, and weak, small, or partially occluded ship targets. The method jointly optimizes image fusion and object detection in one end-to-end model, with explicit feature interaction between the two subtasks: a preliminary fused feature supplements multimodal detection, while object-sensitive detection features are fed back into fusion so that the fused image retains ship semantics more strongly (Guo et al., 28 Jul 2025).

1. Terminology and name disambiguation

The name LSFDNet is unambiguous in the 2025 maritime-imaging paper, where it expands to Long-Short Wave Fusion Detection Network. The surrounding literature, however, contains several similar acronyms that refer to different methods and tasks. This has produced recurring confusion, especially because an earlier super-resolution paper introduced LSFD but not an official architecture called LSFDNet (Guo et al., 28 Jul 2025).

Term	Expansion in the cited paper	Meaning
LSFDNet	Long-Short Wave Fusion Detection Network	Official name of the SWIR–LWIR ship fusion/detection model
LSFD	Local-Selective Feature Distillation	Training/distillation method for SISR, not an official network name
LSDNet	Trainable Modification of LSD Algorithm for Real-Time Line Segment Detection	Unrelated line segment detector
LS-DFN	Dynamic Filtering with Large Sampling Field	Unrelated dynamic filtering module

In particular, “Local-Selective Feature Distillation for Single Image Super-Resolution” does not introduce an official super-resolution architecture named LSFDNet. It introduces LSFD as a teacher-student distillation method applied to standard backbones such as RCAN, EDSR, and SAN, with only the student network used at inference (Park et al., 2021). Similarly, “LSDNet: Trainable Modification of LSD Algorithm for Real-Time Line Segment Detection” is a line detection method rather than a fusion-detection architecture for maritime imaging (Teplyakov et al., 2022), and “Dynamic Filtering with Large Sampling Field for ConvNets” uses the name LS-DFN, not LSFDNet (Wu et al., 2018).

2. Operational setting and sensing rationale

LSFDNet is designed for ship detection in coastal scenes where single-modality imaging is unreliable. The motivating conditions include low illumination or nighttime, haze, fog, smoke, sea-surface clutter such as waves, glare, and highlights, complex nearshore backgrounds, and weak or partially occluded ships. The central premise is that SWIR and LWIR are complementary sensing bands rather than interchangeable ones (Guo et al., 28 Jul 2025).

The paper characterizes SWIR as reflective imaging, roughly 0.9–2.5 $\mu m$ in general, with the dataset camera operating at 0.9–1.7 $\mu m$. SWIR provides better fog and smoke penetration than visible imagery, strong ship-to-sea contrast because seawater absorbs SWIR and appears dark, and richer texture and detail similar to grayscale visible images. Its weakness is dependence on illumination, especially solar irradiance; target brightness decreases under overcast, dusk, or night conditions. By contrast, LWIR, at 8–14 $\mu m$, is thermal imaging. It is robust to illumination changes and supports stable target visibility in low light and at night, but typically lacks fine texture and edge detail and can suffer from thermal diffusion or noisy surface appearance.

This complementarity drives the model design. SWIR contributes texture, contours, and high sea-background contrast; LWIR contributes thermal saliency and illumination robustness. LSFDNet therefore treats fusion and detection as coupled tasks rather than as a conventional cascade in which a fused image is produced first and a detector is trained later. The paper explicitly frames this as an alternative to two-stage pipelines that can prioritize visual quality over downstream utility and limit information exchange between subtasks (Guo et al., 28 Jul 2025).

3. Architecture and cross-task information flow

LSFDNet takes a registered SWIR image $I_{SW}$ and a registered LWIR image $I_{LW}$ as input and consists of three principal components: Multi-Task Feature Extraction (MTFE), the Multi-Level Cross-Fusion (MLCF) branch, and an object detection branch. The architecture is single-stage because it performs feature extraction once in a shared coupled system and trains fusion and detection jointly (Guo et al., 28 Jul 2025).

Within MTFE, the Base Feature Extractor applies three $3\times3$ convolution layers with stride 1 to each modality and outputs 8-channel shallow features $F_{SW}$ and $F_{LW}$. These shallow features are shared by both tasks. The Fusion Feature Extractor then expands channels from 8 to 32 and reduces them back to 8, producing $\tilde F_{SW}$ and $\tilde F_{LW}$. This channel expansion-and-reduction is intended to capture finer-grained features while preserving spatial resolution, which the paper treats as important because fusion is performed at pixel level.

The detection stream uses a YOLOv12 backbone and is described as having four parts: Shallow Feature Extraction, Fusion Feature Augmentation, Deep Feature Extraction, and Multimodal Feature Aggregation. The paper also states that the detector uses A2C2f blocks from YOLOv12, with Area Attention to select important features from multimodal maps. The detection branch receives $\mu m$0, $\mu m$1, and a preliminary fused feature $\mu m$2, and outputs multiscale detection head features $\mu m$3 together with an aggregated detection attention feature $\mu m$4.

The MLCF branch is the fusion core. It is composed of three Multi-Feature Attention (MFA) blocks and performs aggregation across modalities, across scales, and across tasks. At high resolution, MFA-1 fuses $\mu m$5 and $\mu m$6 to produce $\mu m$7. Downsampled features are processed by MFA-2, then upsampled to yield $\mu m$8. MFA-3 combines $\mu m$9 and $\mu m$0 into the preliminary fused feature $\mu m$1. This multiscale design is then extended to multitask fusion by residually injecting $\mu m$2 into the fusion representation, after which a fusion decoder reconstructs the final fused image $\mu m$3 (Guo et al., 28 Jul 2025).

The internal operation of MFA is patch-based. Features are divided into $\mu m$4 patches and vectorized; each modality first undergoes local self-attention, and then bidirectional cross-attention is used so that SWIR queries LWIR thermal information while LWIR queries SWIR structural information. After attention, the patch features are folded back, concatenated, and passed to a local decoder block consisting of four convolution layers, expanding channels from 8 to 16 and reducing them back to 8. The paper’s interpretation is that this design preserves complementary fine and thermal cues, captures information at multiple scales, and introduces object-sensitive semantics from detection into the fusion process.

The information flow is therefore explicitly bidirectional. Fusion $\mu m$5 Detection occurs when $\mu m$6 is sent back into the detection branch through Fusion Feature Augmentation. Detection $\mu m$7 Fusion occurs when $\mu m$8 is residually injected into the fusion branch so that the fused representation focuses more on ships than on sea clutter. The paper presents this as the defining difference between LSFDNet and loose fusion-then-detection coupling (Guo et al., 28 Jul 2025).

4. Objective functions and optimization

LSFDNet is trained with a joint objective

$\mu m$9

where $I_{SW}$0 is the fusion loss and $I_{SW}$1 is the YOLO detection loss function. The paper reports the values $I_{SW}$2, $I_{SW}$3, and $I_{SW}$4, but does not report the numerical value of $I_{SW}$5 (Guo et al., 28 Jul 2025).

A distinctive component is the Object Enhancement (OE) loss, introduced because conventional fusion losses can preserve irrelevant sea-surface clutter as readily as ship content. The fusion loss is

$I_{SW}$6

The global term is

$I_{SW}$7

with

$I_{SW}$8

$I_{SW}$9

and

$I_{LW}$0

Before this global term is defined, the LWIR image is gamma-corrected as

$I_{LW}$1

The paper states that this gamma correction is intended to increase ship-background contrast and suppress sea-surface noise in LWIR (Guo et al., 28 Jul 2025).

The object-focused term is

$I_{LW}$2

where

$I_{LW}$3

$I_{LW}$4

Here $I_{LW}$5 is the number of objects, $I_{LW}$6 indexes object regions, and $I_{LW}$7 are the object-region sizes. The paper emphasizes that the object term uses the original LWIR, not the gamma-corrected LWIR, because ship crops exclude sea clutter and the original thermal information is treated as more trustworthy there. In effect, OE loss splits fusion into global scene fidelity and local ship-focused preservation.

The reported training setup uses Adam, a learning rate of $I_{LW}$8, linear decay, 500 warm-up iterations, 30,000 total iterations, and batch size 8, with NVIDIA GeForce RTX 4090 GPUs and end-to-end joint optimization. The paper does not report separate pretraining, exact data augmentation, optimizer betas, weight decay, mixed precision, FPS, parameter count, FLOPs, or memory use (Guo et al., 28 Jul 2025).

5. Datasets, benchmarks, and empirical behavior

The principal dataset introduced with LSFDNet is NSLSR, the Nearshore Ship Long-Short Wave Registration Dataset, created for SWIR–LWIR fusion and detection. NSLSR contains 1,205 registered SWIR–LWIR image pairs and 2,818 annotated ship objects. The acquisition system uses a SWIR camera with resolution $I_{LW}$9, an uncooled InGaAs FPA detector, $3\times3$0 pixel pitch, and $3\times3$1–$3\times3$2 spectral response, together with an LWIR camera of the same spatial resolution, an uncooled VOx FPA detector, $3\times3$3 pixel pitch, and $3\times3$4–$3\times3$5 spectral response. Registration is performed by manual correction for rigid transformations, additional alignment for soft deformations using MINIMA, and rejection of poorly registered pairs (Guo et al., 28 Jul 2025).

The paper also uses ISD for fusion evaluation. ISD contains 1,044 image pairs at resolution $3\times3$6, but only 28 unique ship instances, with long-range scenes and a monotonous background. For this reason, detection evaluation is performed on NSLSR only. The paper notes an internal inconsistency in the NSLSR split description: the dataset section says 9:1 train/test, whereas the experiments use 844 training images and 361 testing images. Among the NSLSR test set, 118 images are used for fusion evaluation and all 361 for detection evaluation (Guo et al., 28 Jul 2025).

Fusion is evaluated by EN, SF, SD, SCD, VIF, and Q$3\times3$7. On NSLSR, LSFDNet reports EN = 7.181, SF = 21.022, SD = 64.723, SCD = 1.427, VIF = 0.611, and Q$3\times3$8 = 0.520, ranking first on five metrics and second on EN. On ISD, it reports EN = 7.173, SF = 12.330, SD = 50.340, SCD = 1.687, VIF = 0.848, and Q$3\times3$9 = 0.651. The qualitative discussion attributes the method’s advantage to better suppression of sea-surface noise and better preservation of ship details and brightness structure, particularly in cluttered nearshore scenes and difficult lighting (Guo et al., 28 Jul 2025).

Detection is evaluated by precision, recall, mAP$F_{SW}$0, and mAP$F_{SW}$1. The comparison includes single-modal baselines, fused-image detection baselines, and specialized multimodal detectors.

Method	mAP$F_{SW}$2	mAP$F_{SW}$3
SWIR	0.942	0.666
LWIR	0.929	0.628
CAFF-DINO	0.958	0.706
DEYOLO	0.956	0.702
LSFDNet	0.962	0.770

LSFDNet also reports $F_{SW}$4 and $F_{SW}$5. The largest numerical gain is on the stricter localization metric, where it improves over CAFF-DINO from 0.706 to 0.770. The paper further notes a qualitative case in which a small distant ship, partially occluded by the mast of a larger nearby ship, is detected only by LSFDNet (Guo et al., 28 Jul 2025).

The ablation studies isolate the contributions of MLCF, OE loss, and the preliminary fused feature $F_{SW}$6. Without $F_{SW}$7, detection drops to $F_{SW}$8, $F_{SW}$9, $F_{LW}$0, and $F_{LW}$1, versus $F_{LW}$2 and $F_{LW}$3 for the full model. For fusion, the comparison between a plain baseline and the full model shows substantial gains: from EN 6.687, SF 18.365, SD 58.268, SCD 1.023, VIF 0.566, Qabf 0.431 to EN 7.181, SF 21.022, SD 64.723, SCD 1.427, VIF 0.611, Qabf 0.520. The paper explicitly states that MLCF is critical, OE loss helps, multimodal aggregation is especially important, and cross-task interaction helps fusion (Guo et al., 28 Jul 2025).

6. Relation to adjacent work, recurring misconceptions, and limitations

LSFDNet is positioned against two prior paradigms. The first is the two-stage fusion-then-detection pipeline, in which fusion is optimized first and a detector is trained on fused outputs later. The paper argues that such designs can increase training complexity, prioritize image appearance over detection utility, and provide little semantic feedback from detection to fusion. The second is the broader visible–infrared fusion literature. LSFDNet is presented as distinctive because it addresses SWIR + LWIR, not visible + infrared, and because the sensing physics of these bands differ materially: SWIR is reflective and detail-rich but illumination-dependent, while LWIR is emissive, lighting-robust, and texture-poor (Guo et al., 28 Jul 2025).

A frequent misconception is to identify LSFDNet with the super-resolution method LSFD. That is not the terminology used in the super-resolution paper. “Local-Selective Feature Distillation for Single Image Super-Resolution” introduces LSFD as a distillation objective and mechanism inside a teacher-student framework; it explicitly does not introduce an official architecture called LSFDNet. In that work, the underlying networks are standard backbones such as RCAN, EDSR, and SAN, the teacher and distillation modules are discarded at test time, and only the student backbone remains (Park et al., 2021). Likewise, LSFDNet should not be conflated with LSDNet, which is a hybrid CNN-plus-classical detector for line segments (Teplyakov et al., 2022), or with LS-DFN, which is a dynamic filtering module for ConvNets (Wu et al., 2018).

The limitations reported or implied in the LSFDNet paper are domain- and implementation-specific. The method assumes registered SWIR/LWIR input, and the authors note that the dataset itself required manual correction and additional registration; performance may therefore degrade when misalignment is significant. The model is tuned for maritime ship detection and sea-background suppression, so generalization beyond that domain is not established. The paper omits several implementation specifics and does not report runtime, parameter count, FLOPs, or memory use. NSLSR fills an important data gap, but its scale remains moderate relative to large object-detection datasets, and the paper contains a split inconsistency between the stated 9:1 train/test ratio and the explicit 844/361 counts used experimentally (Guo et al., 28 Jul 2025).