DELNet: Dynamic Expert Neural Networks

• DELNet is a framework of dynamic expert neural networks tailored for multi-domain tasks like weather restoration, ECG delineation, and mobile ISP.
• It employs specialized modules such as a judging valve, dynamic expert library, and attention mechanisms to efficiently route and process diverse inputs.
• DELNet’s design minimizes retraining and computational costs while achieving high performance across image, signal, and mobile processing benchmarks.

DELNet is an acronym that denotes distinct neural network architectures applied across multiple technical domains. This entry surveys the most prominent formulations, focusing on (1) continual all-in-one weather image restoration, (2) end-to-end electrocardiogram (ECG) delineation (often “ECG-DelNet”), and (3) single-stage mobile image signal processing (ISP), each leveraging specialized adaptations of deep learning principles.

1. Core Definitions and Problem Scope

DELNet (Dynamic Expert Library Network) as formulated for weather image restoration is a continual learning framework designed to handle diverse image degradations by dynamically managing a bank of expert networks. The model integrates a “judging valve” to differentiate between seen and novel restoration tasks and a dynamic expert library (“adapter pool”) for specialized processing and knowledge transfer, minimizing retraining overhead and catastrophic forgetting (Liu et al., 30 Jan 2026).

In ECG delineation (“ECG-DelNet”), DELNet refers to a U-Net-derived, one-dimensional convolutional architecture optimized for segmenting ECG signals into clinically relevant regions (P wave, QRS complex, T wave) (Jimenez-Perez et al., 2020).

The term is also used for “Del-Net,” a single-stage deep convolutional network that replaces conventional cascaded modules in mobile camera ISPs, integrating spatial and channel attention with lightweight residual attention blocks for efficient Bayer-to-sRGB image conversion (Gupta et al., 2021).

2. Architecture and Methodological Innovations

2.1. DELNet for Weather Restoration

DELNet’s architecture comprises three primary modules (Liu et al., 30 Jan 2026):

• Deep Feature Enhancement (DFE) Backbone: Enhances multi-scale spatial-channel features using parallel self-attention and polarization attention, soft-max pooling, and weight mapping.
• Judging Valve (JV): Computes a task vector $T^t$ from encoder features using channel-wise statistics (mean, standard deviation, max, min, Euclidean norm). Task similarity across sessions is quantified using cosine, Euclidean, and Pearson metrics, combined into a unified score $S_{sum}$. This mechanism autonomously routes inputs to the most appropriate set of experts or triggers adaptation for novel weather conditions.
• Dynamic Expert Library (DEL): Maintains an expandable pool of lightweight adapters. Each expert is a stack of instance normalization, $1 \times 1$ convolutions, GELU activation, and a residual pathway. Experts are activated, trained, and frozen according to task similarity and adaptive scoring based on both historical performance and usage statistics.

The selection, fusion, and update steps are mathematically formalized. Expert outputs are fused using temperature-scaled soft weighting, with the total loss comprising multi-level terms: image reconstruction, contrastive alignment, knowledge distillation, feature-level projection, regularization, and expert diversity (Liu et al., 30 Jan 2026).

2.2. ECG-DelNet for Signal Delineation

The ECG-DelNet adapts the canonical U-Net to one-dimensional signal segmentation (Jimenez-Perez et al., 2020):

• Encoder–Decoder Structure: Five downsampling/upsampling stages, with feature channels doubling and halving (32→512, bottleneck at 1024, reverting to 32). Downsampling uses 1D average pooling, upsampling relies on bilinear interpolation and $1 \times 1$ convolution.
• Convolutional Blocks: Each consists of two consecutive 1D convolutions (kernel size 3), ReLU, batch normalization, and optional spatial dropout ($p = 0.25$).
• Skip Connections: Concatenation of encoder and decoder tensors at matching temporal resolutions maximizes information flow for fine-grained wave delineation.
• Loss Function: Segmentation is supervised using a soft-Jaccard (IoU) loss summed over foreground classes ($\text{P, QRS, T}$), which directly optimizes for overlap between predicted and ground truth regions.

2.3. Del-Net for Mobile ISP

Del-Net discards the hand-crafted modular approach, instead fusing multi-scale and single-scale branches (Gupta et al., 2021):

• “Flat + Pyramidal” Design: A full-resolution stream of Enhanced Attention Modules (EAM) complements a U-Net-style encoder–decoder, with Spatial-and-Channel Attention (SCA) blocks at each scale.
• SCA Blocks: Combine global channel attention (via pooled descriptors, two-layer MLP, sigmoid gating) and spatial attention (max/avg pooled maps, 7×7 convolution, sigmoid), merged additively with a residual connection.
• EAM blocks: Apply parallel dilated convolutions to capture multi-scale features, fuse outputs additively with the input (skip connection), and finalize with channel attention gating.
• Output Reconstruction: Aggregation of both branches feeds a lightweight pixel-shuffling head for final sRGB restoration.

3. Training Paradigms and Regularization Approaches

3.1. DELNet (Weather Restoration)

Each new degradation type triggers an adaptive update cycle governed by the judging valve: matched experts are reused; otherwise, top-K experts are activated, and a new adapter is instantiated and optimized. Only the relevant experts/adapters and the DFE are updated at each task, reducing compute demand. Regularization is enforced through $L_2$ penalties and explicit diversity-promoting terms (Liu et al., 30 Jan 2026).

3.2. ECG-DelNet

To address data scarcity, DELNet is pre-trained on low-quality (automatic DSP) annotations from the full QT database, then fine-tuned on high-quality expert-annotated segments. On-the-fly data augmentation simulates AWGN, random and pacemaker spikes, amplitude saturation, and frequency-specific noise to enhance generalizability. Regularization exploits spatial dropout, batch normalization, and $L_2$ weight decay (Jimenez-Perez et al., 2020).

3.3. Del-Net (Mobile ISP)

Del-Net is trained end-to-end using a weighted combination of $L_1$ (including log difference), perceptual, and MS-SSIM losses. Data augmentation is limited to basic geometric and normalization transforms (Gupta et al., 2021).

4. Quantitative Performance and Empirical Findings

Weather Restoration

DELNet achieves an average PSNR/SSIM of 31.28 dB/0.94 across the RESIDE (OTS haze), Rain100H, and Snow100K benchmarks, surpassing prior continual-learning baselines (AFC at 25.73/0.77) and even outperforming static all-in-one methods such as CLAIO (30.67/0.93). Gains are distributed across all task types and persist in both sequential and multi-task settings. DELNet achieves these results with only 5.6 M parameters, a marked reduction compared to TransWeather (38 M) and WeatherDiff (83 M) (Liu et al., 30 Jan 2026).

ECG Delineation

The single-lead configuration delivers precision and recall as follows (5-fold cross-validation):

• P-wave: Precision = 90.12%, Recall = 98.73%
• QRS complex: Precision = 99.14%, Recall = 99.94%
• T-wave: Precision = 98.25%, Recall = 99.88%

Perfomance is on par with DSP-based approaches, but with greatly reduced hand-tuning and real-time CPU inference (<10 ms/window) (Jimenez-Perez et al., 2020).

Mobile ISP

Del-Net demonstrates competitive RAW-to-sRGB performance (PSNR ≈ 21.5 dB, SSIM ≈ 0.75) on the Zurich dataset, with computational cost reduced by ~90%: 2.68 M parameters, 0.53 × 10¹² Mult-Adds (versus 47.6 M parameters and 20.3 × 10¹² Mult-Adds for PyNET). Ablations confirm the contribution of EAM and SCA blocks (Gupta et al., 2021).

Model	Params (M)	Mult-Adds (×10¹²)	PSNR (dB)	SSIM	CIEDE2000
PyNET	47.6	20.3	21.19	0.746	9.99
AWNet	96.1	10.2	21.86	0.780	9.76
MW-ISPNet	29.2	5.2	21.91	0.784	9.71
Del-Net	2.68	0.53	21.46	0.745	9.64

5. Limitations and Open Challenges

For weather restoration, DELNet’s final performance is mildly sensitive to task order; optimal expert pool size depends on the trade-off between specialization and memory footprint. Dynamic thresholding in the judging valve requires careful tuning for nonstationary degradation sequences. For highly mixed degradations, expert combinatorics may necessitate more scalable routing strategies (Liu et al., 30 Jan 2026).

In ECG delineation, data hunger and ground-truth inconsistencies—especially at T-onsets—introduce error bounds that limit achievable accuracy, and model interpretability remains constrained relative to hand-engineered DSP approaches (Jimenez-Perez et al., 2020).

In mobile ISP, while Del-Net reduces computational load dramatically, absolute image quality margins relative to the most computationally intensive models (e.g., MW-ISPNet) are narrow (Gupta et al., 2021).

6. Practical Deployment and Impact

The DELNet framework for continual all-in-one weather removal is engineered for real-time deployment, requiring only the storage and periodic update of adapter weights and task vectors. The architecture is well matched to autonomous and surveillance applications where new weather events must be processed without retraining or performance loss on prior types (Liu et al., 30 Jan 2026).

ECG-DelNet’s end-to-end design enables rapid, robust delineation suitable for ambulatory and wearable cardiac monitoring devices, obviating manual rule-tuning and supporting downstream tasks such as arrhythmia screening and interval analysis (Jimenez-Perez et al., 2020).

Del-Net’s efficiency renders it compatible with computational and memory restrictions of contemporary mobile devices, permitting replacement of modular ISPs with a single, differentiable network (Gupta et al., 2021).

---

DELNet designates specialized neural architectures tailored to demanding real-world tasks—continual weather image restoration, precise 1D biomedical signal segmentation, and resource-efficient mobile ISP—each exemplifying distinct approaches to modularity, attention, and regularization to address the practical and theoretical challenges of their domains.