Dual Perception Network (DPNet)

• DPNet is a unified framework that jointly enhances underwater images and detects objects using a shared representation.
• It employs a bilevel optimization strategy to coordinate task-specific subnets, improving both enhancement quality and detection accuracy.
• DPNet’s design enables selective inference, allowing only the necessary subnet to be used at test time to reduce computational overhead.

Searching arXiv for the primary DPNet paper and closely related uses of the acronym. Dual Perception Network (DPNet) most specifically denotes a neural architecture introduced for jointly learning underwater image enhancement and object detection through a bilevel optimization formulation that is unrolled into a network with one shared module and two task subnets (Fu et al., 2023). In that usage, DPNet seeks a shared representation that provides more structural details for image enhancement and rich content information for object detection, and it is optimized with a cooperative training strategy (Fu et al., 2023). The acronym is, however, context-dependent in the literature: closely related or homonymous forms include DP-Net for simultaneous individual and cooperative perception in connected and automated vehicles (Qu et al., 2023), DPNet for efficient or real-time object detection with lightweight attention (Shi et al., 2021, Zhou et al., 2022), and DPNet for deepfake cross-domain detection (Jia et al., 14 Aug 2025). This ambiguity makes the task domain essential when interpreting the term.

1. Definition and scope

In the underwater vision literature, DPNet is a unified framework for two interrelated tasks: underwater object detection and image enhancement (Fu et al., 2023). The motivating premise is that existing approaches either improve feature representation inside detectors without eliminating degradation in image content such as color and texture, or generate visually appealing enhanced outputs that do not necessarily yield high detector accuracy; multi-task alternatives can improve both tasks but often invoke huge architecture and expensive computations (Fu et al., 2023). DPNet addresses this by jointly learning the two tasks through a shared representation and a bilevel optimization scheme (Fu et al., 2023).

A common misconception is to treat “DPNet” as a single, universally recognized architecture. The literature instead uses the label for several unrelated designs. In underwater perception, DPNet denotes a joint enhancement–detection model (Fu et al., 2023). In cooperative vehicular perception, DP-Net denotes a lightweight feature processor that supports both individual and cooperative 3D detection (Qu et al., 2023). In lightweight object detection, DPNet denotes a dual-path detector with lightweight self-attention or lightweight self-correlation modules (Shi et al., 2021, Zhou et al., 2022). In deepfake detection, DPNet denotes a dual-pathway perception backbone composed of a frequency stream and a graph-convolution-based relationship module (Jia et al., 14 Aug 2025). The shared acronym therefore identifies a design pattern of dual or joint perception rather than a single canonical model.

2. Underwater DPNet architecture

The underwater DPNet consists of three main modules: a shared module, an image enhancement subnet, and an object detection subnet (Fu et al., 2023). The shared module extracts a rich shared representation from the input image, providing both content such as color and texture and structure information (Fu et al., 2023). The enhancement subnet takes the shared features and outputs a high-quality, enhanced version of the underwater image, while the detection subnet utilizes the shared representation for robust object detection (Fu et al., 2023).

The shared module is lightweight and suitable for mobile deployment; it contains 3 convolutional layers with kernels $3 \times 3$, $5 \times 5$, and $3 \times 3$, and channel size 32 (Fu et al., 2023). Its role is to extract and transmit low-level information, including color, texture, and structure (Fu et al., 2023). The enhancement subnet uses 3 plain $3 \times 3$ convolutional layers, without downsampling or upsampling, avoiding information loss and architectural bloat (Fu et al., 2023). It uses BatchNorm and activation functions, with Relu for early layers and Sigmoid for output ensuring results in $[0,1]$, and outputs an RGB enhanced image (Fu et al., 2023). The detection subnet is based on RetinaNet, with Focal Loss for class balance and FPN for multi-scale robust feature extraction; its first-layer channel size is aligned with the shared module at 32 (Fu et al., 2023).

During training, all modules are active; at inference, only the shared module and one task-specific subnet is used (Fu et al., 2023). This truncation property is central to the framework’s deployment logic. A plausible implication is that DPNet treats joint training primarily as a mechanism for representation learning, while allowing task-specific inference paths that avoid permanently carrying the full multi-task architecture at test time.

3. Bilevel optimization and cooperative training

The authors formulate joint training as a bilevel optimization problem inspired by Stackelberg’s theory (Fu et al., 2023). The objective is written as

$$\begin{aligned} &\min_{\bm u}~\mathcal{L}^{val}(\bm{\omega}^*, \Phi (\bm{u}))\ &\text{s.t. }~\bm{\omega}^* = \arg\min_{\bm{\omega}}\, \mathcal{L}^{tr}(\bm{\omega}, \Phi (\bm{u})) \end{aligned}$$

where $\Phi$ is the shared module parameterized by $\bm u$, $\bm{\omega}=(\bm{\omega}_{det},\bm{\omega}_{enh})$ are detection and enhancement subnet parameters, and $$\mathcal{L} = \mathcal{L}_{det} + \mathcal{L}_{enh}$$ is the joint loss for both tasks (Fu et al., 2023). $5 \times 5$0 and $5 \times 5$1 denote validation and training losses, respectively (Fu et al., 2023).

Within this structure, shared parameters $5 \times 5$2 are updated in the outer loop, while task subnet parameters $5 \times 5$3 are updated in the inner loop given shared module outputs (Fu et al., 2023). The stated purpose is to force the shared module to learn representations useful for both enhancement and detection (Fu et al., 2023). The parameter gradient is given as

$5 \times 5$4

which unifies updates for the shared module and task-specific parameters (Fu et al., 2023).

The cooperative training strategy follows directly from the bilevel scheme. The shared module parameters are updated considering both detection and enhancement validation loss, while each task subnet is independently optimized for its own loss in the lower level given the current shared representation (Fu et al., 2023). Implementation details reported for this setting are Adam optimizer, 24 epochs, and batch size 2 (Fu et al., 2023). After training, the shared module learns features boosting both detection and enhancement, and at inference only the required subnet is kept with the shared module, reducing compute and memory (Fu et al., 2023).

4. Shared representation and task interaction

The central representation claim is that the shared representation provides more structural details for image enhancement and rich content information for object detection (Fu et al., 2023). For detection, the enhanced content extracted by the shared module helps in challenging underwater scenes where color cast or low contrast obscures objects (Fu et al., 2023). For enhancement, structural information passed along for detection provides cues to produce sharper, more realistic enhanced images, including crisp boundaries and preserved details (Fu et al., 2023).

This mutual positive feedback is presented as the explanation for the model’s gains on both tasks (Fu et al., 2023). A plausible implication is that DPNet differs from serial enhancement-then-detection pipelines by not treating enhancement as a pre-processing stage with a fixed visual objective. Instead, enhancement and detection are coupled through a representation that is optimized under both objectives. The paper explicitly notes that visually appealing enhancement modules do not necessarily generate high accuracy for deep detectors, which is why the joint shared representation is positioned as necessary rather than optional (Fu et al., 2023).

The architecture is also framed as efficient relative to prior multi-task systems. Most prior joint methods, such as HybridDGAN, rely on heavy architectures with both enhancement and detection fully active at inference, leading to large size and inefficiency (Fu et al., 2023). By contrast, DPNet can be truncated for “detection only” or “enhancement only” use with minimum compute at test time (Fu et al., 2023). This design choice places the shared module, rather than late fusion or full co-execution, at the center of the method.

5. Experimental results in underwater enhancement and detection

Experiments are reported on synthetic Underwater-Cityscapes and real UEDB datasets for detection, with comparisons against general detectors such as RetinaNet and YOLOX, underwater-specific detectors such as AquaNet, CSAM, and FERNet, enhancement+detection pipelines such as FUnIE-GAN/ERH plus detection models, and other joint or multitask methods such as HybridDGAN (Fu et al., 2023). On UEDB, DPNet achieves $5 \times 5$5 mAP, outperforming HybridDGAN at $5 \times 5$6, FERNet at $5 \times 5$7, and RetinaNet at $5 \times 5$8 (Fu et al., 2023). On Underwater-Cityscapes, DPNet achieves $5 \times 5$9 mAP (Fu et al., 2023). Qualitatively, it is reported as robust under haze, color cast, and light interference (Fu et al., 2023).

For enhancement, the reported metrics are SSIM, PSNR, PCQI, UIQM, and NIQE, with baselines ERH, FUnIE-GAN, HybridDGAN, and UDCP (Fu et al., 2023). DPNet consistently achieves the highest or near-highest scores on enhancement metrics across both datasets, and outputs are described as visually appealing and more natural (Fu et al., 2023).

The paper also reports efficiency comparisons. For detection, DPNet uses 70.3M parameters and 4.3 FPS, compared with HybridDGAN at 148.7M parameters and 2.4 FPS (Fu et al., 2023). For enhancement, DPNet uses 30.2M parameters and 5.5 FPS (Fu et al., 2023). The key efficiency claim is that, because only the shared module and one subnetwork are needed at inference, DPNet yields significant space and computation savings relative to prior multi-task models that always keep both branches active (Fu et al., 2023).

Setting	DPNet result	Reported comparison
UEDB detection	43.7% mAP	HybridDGAN 39.8%, FERNet 36.3%, RetinaNet 34.0%
Underwater-Cityscapes detection	39.8% mAP	Similar clear gaps over baselines
Detection efficiency	70.3M params, 4.3 FPS	HybridDGAN 148.7M params, 2.4 FPS
Enhancement efficiency	30.2M params, 5.5 FPS	Shared module plus one subnetwork at inference

These numbers define DPNet’s empirical profile in its original underwater setting: a jointly trained but selectively deployable system that improves both tasks while reducing inference burden relative to heavier joint models (Fu et al., 2023).

6. Related architectures sharing the DPNet or dual-perception label

The term “dual perception” recurs across several research areas, but with substantially different operational meanings.

In connected and automated vehicles, SiCP introduces a Dual-Perception Network (DP-Net) as a generic framework component for both individual and cooperative perception (Qu et al., 2023). There, DP-Net is a lightweight feature processor with only 0.13M parameters that forwards ego features in standalone mode and fuses transformed neighbor feature maps in cooperative mode using a learned weighted fusion map $3 \times 3$0 intended to retain crucial gradient information during feature map fusion (Qu et al., 2023). On OPV2V “Default,” SiCP reports $3 \times 3$1 [email protected] and $3 \times 3$2 [email protected] for cooperative perception, while in individual mode it reports $3 \times 3$3 [email protected] and $3 \times 3$4 [email protected], matching or slightly surpassing PointPillars without degradation in solo mode (Qu et al., 2023). Here, “dual perception” denotes simultaneous support for individual and cooperative perception rather than joint optimization of two vision tasks.

In lightweight object detection, two different papers use DPNet for a dual-path network. One presents DPNet as an efficient object detector with a lightweight self-attention module (LSAM) and a lightweight cross-attention module (LCAM) in the FPN, reporting $3 \times 3$5 AP on COCO test-dev with 1.14 GFLOPs and 2.27M parameters for $3 \times 3$6 input (Shi et al., 2021). Another presents DPNet as a real-time dual-path detector with a lightweight self-correlation module (LSCM) and a lightweight cross-correlation module (LCCM), reporting $3 \times 3$7 AP on MS COCO test-dev, $3 \times 3$8 mAP on Pascal VOC 2007 test set, nearly 2.5M model size, 1.04 GFLOPs, and 164 FPS and 196 FPS for $3 \times 3$9 input images of the two datasets (Zhou et al., 2022). In both cases, the “dual” structure refers to parallel high-resolution and low-resolution paths for localization and semantics, not to multi-task shared representations.

In deepfake cross-domain detection, a 2025 paper defines DPNet as part of a Forgery Guided Learning strategy (Jia et al., 14 Aug 2025). That DPNet comprises a Frequency-domain Perception Mechanism and an Adaptive Forgery Relationship Perception module, combining a frequency stream, a spatial stream, and graph convolution to perceive both differences and relationships among forgery traces (Jia et al., 14 Aug 2025). It is supervised with binary cross-entropy and intra-class compact loss, and is reported to achieve AUC $3 \times 3$0 on FF++ and average AUC $3 \times 3$1 on cross-dataset evaluation (Jia et al., 14 Aug 2025). The acronym thus travels across domains, but its internal semantics shift with the task.

7. Conceptual significance and interpretive issues

Across its various instantiations, DPNet typically indicates an architectural attempt to combine complementary perceptual sources: enhancement and detection in underwater imagery (Fu et al., 2023), individual and cooperative sensing in vehicular perception (Qu et al., 2023), high-level semantics and low-level spatial details in lightweight detection (Shi et al., 2021, Zhou et al., 2022), or frequency and relational cues in deepfake detection (Jia et al., 14 Aug 2025). This suggests a family resemblance centered on pairing distinct information channels rather than a single fixed blueprint.

Within the underwater formulation, the most significant design decision is the use of bilevel optimization to organize multi-task learning (Fu et al., 2023). The paper’s factual claims indicate that the shared representation is not merely a parameter-sharing convenience but the explicit optimization target through which both tasks are coordinated (Fu et al., 2023). A plausible implication is that the architecture belongs as much to the literature on task-coupled representation learning as to underwater enhancement or detection alone.

The principal source of confusion is nomenclatural rather than technical. “DPNet” may refer to at least three distinct object-detection architectures, one underwater joint learning model, one cooperative vehicular fusion module, and one deepfake detector (Fu et al., 2023, Shi et al., 2021, Zhou et al., 2022, Qu et al., 2023, Jia et al., 14 Aug 2025). For research usage, the paper title and domain are therefore indispensable identifiers. In strict encyclopedic usage, the underwater model introduced in “Joint Perceptual Learning for Enhancement and Object Detection in Underwater Scenarios” is the most direct referent of “Dual Perception Network (DPNet)” (Fu et al., 2023), while the broader literature shows that the label has evolved into a recurrent naming pattern for architectures built on complementary perceptual pathways.