Deep Evidential Fusion Network (DEFNet) Overview

• Deep Evidential Fusion Network (DEFNet) is a deep learning paradigm that fuses multiple predictive evidences while retaining explicit uncertainty through methods like Dirichlet, Dempster–Shafer, or NIG formulations.
• It supports diverse applications—such as blind image quality assessment, medical segmentation, remote sensing, and stereo matching—by working effectively with misaligned, noisy, or incomplete data.
• Advanced fusion operators and reliability mechanisms within DEFNet enhance decision confidence by managing conflicts and calibrating uncertainty, thereby improving model accuracy and robustness.

Deep Evidential Fusion Network (DEFNet) denotes, across recent arXiv literature, a family of deep architectures that fuse multiple predictive sources while preserving explicit uncertainty in an evidential representation. The label appears explicitly in the blind image quality assessment model "DEFNet: Multitasks-based Deep Evidential Fusion Network for Blind Image Quality Assessment" (Lou et al., 25 Jul 2025). Closely related works state that their methods are not explicitly named DEFNet but can be understood as, or are an apt shorthand for, a DEFNet-style system, including multi-vehicle evidential occupancy-grid fusion (Kempen et al., 2023), semi-supervised medical segmentation (He et al., 2024), aerial-ground scene classification (Zhao et al., 2023), and incomplete multi-view classification with conflict-aware discount fusion (Huang et al., 2024). Across these instantiations, DEFNet replaces purely point-estimate fusion with belief masses, ignorance terms, Dirichlet parameters, or normal–inverse-gamma (NIG) parameters, and couples those representations to fusion operators designed to remain informative under conflict, misalignment, missing views, or pseudo-label noise.

1. Terminology, scope, and canonical usage

Across the cited works, DEFNet is used in two distinct but compatible ways. In BIQA, it is the formal name of a multitask architecture (Lou et al., 25 Jul 2025). In several other papers, the same label is introduced only as a conceptual shorthand: the occupancy-grid fusion model in live digital twins is described as a "deep evidential fusion network in spirit" (Kempen et al., 2023); the semi-supervised medical segmentation framework built from IPAF and VWAL is said to admit DEFNet as an apt shorthand (He et al., 2024); the aerial-ground scene classifier is named Evidential Fusion Network (EFN) but is functionally described as DEFNet-style (Zhao et al., 2023); and the incomplete multi-view classifier instantiates DEFNet via its Conflict-Aware Evidential Fusion Network (CAEFN) (Huang et al., 2024).

This suggests that DEFNet functions less as a single fixed architecture than as a recurring evidential-fusion pattern. In that pattern, multiple views, modalities, branches, regions, or predictors generate evidential outputs; these outputs are aligned on a common evidential semantics; and a fusion module combines them while retaining an explicit uncertainty term rather than collapsing everything into softmax scores.

A recurrent misconception is to equate DEFNet exclusively with Dirichlet-based Evidential Deep Learning. The literature is broader. Some variants are Dirichlet/Subjective-Logic systems, some are prototype-based Dempster–Shafer networks, and some use NIG evidential regression rather than categorical evidence.

2. Evidential representations and uncertainty semantics

A large subset of DEFNet variants uses the Dirichlet/Subjective-Logic parameterization. For a $K$-class problem, a network predicts nonnegative evidence $e_k$, defines $\alpha_k=e_k+1$, computes total strength $S=\sum_{k=1}^K \alpha_k$, and derives belief and uncertainty by

$$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$

This formulation appears in evidential occupancy-grid fusion, semi-supervised medical segmentation, dual-view remote sensing classification, incomplete multi-view classification, and multiview PolSAR classification (Kempen et al., 2023, He et al., 2024, Zhao et al., 2023, Huang et al., 2024, Shi et al., 13 Oct 2025). In these models, large $S$ implies high confidence, while small $S$ increases the ignorance mass $u$.

Several DEFNet variants instead operate directly with Dempster–Shafer basic belief assignments. In multimodal medical segmentation with contextual discounting, each modality-specific evidential head outputs a mass function $m_n^t$ whose focal sets are the singleton classes and $\Theta$, where $e_k$0 is explicit ignorance (Huang et al., 2023). In deep evidential PET/CT lymphoma segmentation, the evidential layer maps prototype distances in feature space to masses $e_k$1, $e_k$2, and $e_k$3, and the pignistic probability of lymphoma is

$e_k$4

That model therefore realizes uncertainty through feature-space support rather than through a Dirichlet concentration alone (Huang et al., 2022).

A third line uses evidential regression. In stereo matching and BIQA, the network predicts NIG parameters $e_k$5 or $e_k$6, where $e_k$7 is the predictive mean and the remaining parameters encode uncertainty over the mean and variance (Lou et al., 2023, Lou et al., 25 Jul 2025). In stereo, the derived quantities are reported explicitly as

$e_k$8

so aleatoric and epistemic components are both represented (Lou et al., 2023).

When a classical probability is needed for decision making, many DEFNet variants apply a pignistic transform. In binary occupancy-grid fusion, for instance,

$e_k$9

which redistributes ignorance evenly between free and occupied hypotheses (Kempen et al., 2023).

3. Fusion operators and reliability mechanisms

The best-known fusion rule in DEFNet-related work is Dempster’s Rule of Combination. In evidential occupancy-grid fusion, the classical baseline transforms one OGM into the other frame and fuses cellwise by

$\alpha_k=e_k+1$0

but the paper notes that this baseline is highly sensitive to spatial misalignment and amplifies conflicts rather than correcting alignment (Kempen et al., 2023). The same reliance on DS fusion appears in multiview evidential image classification and PolSAR fusion, where evidence from different classifiers or manifold branches is combined on a common frame (Tong et al., 2021, Shi et al., 13 Oct 2025).

Because direct DS normalization can become brittle under conflict, many DEFNet variants introduce conservative or reliability-aware alternatives. IPAF in semi-supervised medical image segmentation fuses singleton masses by

$\alpha_k=e_k+1$1

while retaining

$\alpha_k=e_k+1$2

The coefficient $\alpha_k=e_k+1$3 limits conversion of uncertainty into singleton confidence and explicitly avoids the overconfident normalization of classical Dempster fusion (He et al., 2024).

EFN for aerial-ground classification uses a different risk-aware operator. Each view provides an opinion $\alpha_k=e_k+1$4, and fusion produces

$\alpha_k=e_k+1$5

with $\alpha_k=e_k+1$6 enforcing normalization. This operator downweights high-risk views rather than treating views symmetrically (Zhao et al., 2023).

CAEFN in incomplete multi-view classification introduces learnable view-specific discount factors $\alpha_k=e_k+1$7. Its conflict-aware aggregation is proved equivalent to weighted evidence pooling,

$\alpha_k=e_k+1$8

so reliability is learned end-to-end rather than fixed a priori (Huang et al., 2024). Multimodal medical segmentation with contextual discounting adopts class- and modality-specific reliabilities $\alpha_k=e_k+1$9 and discounts each modality’s contour by

$S=\sum_{k=1}^K \alpha_k$0

before combining discounted contours multiplicatively across modalities (Huang et al., 2023).

NIG-based DEFNet variants use additive evidential fusion. In BIQA, the operator

$S=\sum_{k=1}^K \alpha_k$1

$S=\sum_{k=1}^K \alpha_k$2

fuses local and global evidential regressors in closed form (Lou et al., 25 Jul 2025). Stereo matching uses the same mixture-of-NIG principle for both intra-branch and inter-branch fusion (Lou et al., 2023).

4. Architectural patterns

DEFNet architectures are diverse, but several motifs recur. One motif is channelwise or tensorwise concatenation followed by a shared evidential backbone. The occupancy-grid model concatenates two prealigned OGMs into a 4-channel tensor $S=\sum_{k=1}^K \alpha_k$3 and processes it with DeepLabV3+ using a ResNet-50 backbone, ASPP, and a decoder whose final activation is changed from softmax to ReLU so that the network outputs nonnegative evidence (Kempen et al., 2023).

A second motif is the dual-stream or multi-stream design. EFN for remote sensing assigns an independent backbone to each view, with backbones such as AlexNet, VGG-11, ResNet-18, Inception, and DenseNet, and replaces the terminal softmax with a non-negative evidential head (Zhao et al., 2023). CAEFN in partial multi-view classification uses view-specific DNNs $S=\sum_{k=1}^K \alpha_k$4, each with an evidence bottleneck, followed by a weighted pooling fusion layer $S=\sum_{k=1}^K \alpha_k$5 parameterized by learnable discount factors (Huang et al., 2024).

A third motif is geometry- or modality-specific branching. MMEFnet for PolSAR constructs one branch on the HPD manifold and another on the Grassmann manifold, with manifold-aware sGCNs and an evidential classifier replacing the conventional softmax head (Shi et al., 13 Oct 2025). ELFNet for stereo matching divides the problem into a local cost-volume branch and a global transformer branch, each terminating in a trustworthy head that emits valid NIG parameters (Lou et al., 2023). The BIQA DEFNet uses CLIP with ViT-B/32 and GPT-2, processes four local crops during training and a global downsampled image, and fuses region-level and local-global NIGs (Lou et al., 25 Jul 2025).

Medical segmentation variants tend to use encoder-decoder backbones. IPAF+VWAL employs 3D V-Net for LA, Pancreas-CT, and TBAD and 2D U-Net for ACDC (He et al., 2024). Contextual-discounting DEFNet is explicitly late-fusion and decision-level: each modality has its own feature extractor and evidential module, implemented with UNet, nnUNet, or nnFormer backbones (Huang et al., 2023). Deep evidential PET/CT lymphoma segmentation uses a shared 3D residual U-Net followed by a prototype-based evidential layer rather than a softmax classifier (Huang et al., 2022).

These designs show that DEFNet is architecture-agnostic at the backbone level. The common structural requirement is not a specific encoder, but the insertion of a head that produces evidential quantities and a downstream fusion module that manipulates those quantities directly.

5. Objectives, supervision, and curricula

Loss design is as central to DEFNet as architecture. In the occupancy-grid model, the per-cell loss combines squared error on the Dirichlet means with a variance term,

$S=\sum_{k=1}^K \alpha_k$6

with additional weighting for occupied cells because they are underrepresented (Kempen et al., 2023). EFN introduces a Reciprocal Loss in which the positive-class term is the Bayes risk of cross-entropy under the Dirichlet and the negative-class term penalizes spurious evidence through a reciprocal factor involving the digamma function (Zhao et al., 2023).

Several multiview classification DEFNet variants adopt the standard EDL objective based on Dirichlet-integrated cross-entropy plus a KL term to a non-informative prior. CAEFN optimizes $S=\sum_{k=1}^K \alpha_k$7, with annealing $S=\sum_{k=1}^K \alpha_k$8, and adds an auxiliary sum over per-view losses (Huang et al., 2024). MMEFnet uses the same digamma-based evidential cross-entropy and KL regularizer over both fused and per-view Dirichlet parameters (Shi et al., 13 Oct 2025).

NIG-based regression variants optimize negative log-model-evidence together with an evidence regularizer. In stereo matching, the uncertainty loss is

$S=\sum_{k=1}^K \alpha_k$9

where $$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$0 is the NIG negative log-evidence and $$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$1 discourages unwarranted evidence (Lou et al., 2023). BIQA adopts the same evidential principle on top of multitask quality, scene, and distortion objectives, and its global objective is

$$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$2

where $$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$3 and $$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$4 correspond to cross sub-region and local-global evidential fusion losses (Lou et al., 25 Jul 2025).

Semi-supervised medical segmentation adds curriculum mechanisms. IPAF+VWAL defines

$$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$5

and ranks voxels by this fused uncertainty so that the weight

$$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$6

progressively shifts learning toward harder voxels (He et al., 2024). Co-evidential fusion with IVUM augments evidential learning with Information Volume of Mass Function and introduces IVUM-weighted objectives (He et al., 3 Jun 2025). MEDL further combines class-aware evidential fusion with an asymptotic Fisher-information-based evidential loss and a reliability mask derived from fused uncertainty and entropy (He et al., 18 May 2025).

6. Domains, benchmarks, and empirical behavior

The empirical record of DEFNet-style methods is unusually broad, spanning traffic digital twins, medical segmentation, remote sensing, partial multi-view classification, PolSAR, stereo matching, and BIQA.

Domain	Representative system	Reported outcome
Live traffic digital twins	OGM fusion (Kempen et al., 2023)	Compensates misalignments up to $$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$7 m and $$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$8
Semi-supervised medical segmentation	IPAF + VWAL (He et al., 2024)	LA, 5% labels: Dice 90.50%, Jaccard 82.76%
Aerial-ground remote sensing	EFN (Zhao et al., 2023)	AiRound, VGG-11: product 90.41% vs evidential fusion 92.16%
Incomplete multi-view classification	CAEFN (Huang et al., 2024)	ROSMAP misaligned $$b_k=\frac{e_k}{S}, \qquad u=\frac{K}{S}, \qquad p_k=\frac{\alpha_k}{S}.$$9: 0.6928 vs UIMC 0.5857
PolSAR classification	MMEFnet (Shi et al., 13 Oct 2025)	Flevoland: OA 99.75, AA 99.40, Kappa 99.73
Stereo matching	ELFNet (Lou et al., 2023)	Scene Flow: EPE 0.33, D1-1px 1.28
Blind image quality assessment	DEFNet (Lou et al., 25 Jul 2025)	KADID-10k: SRCC 0.942, PLCC 0.944

Beyond these representative figures, related papers report gains in uncertainty quality or reliability. In PET/CT lymphoma segmentation, contextual-discounting multimodal evidential fusion improves Dice from 0.770 to 0.811 relative to UNet and reduces ECE, Brier, and NLL (Huang et al., 2023). In deep evidential PET/CT lymphoma segmentation with a prototype-based evidential layer, ENN-UNet reaches Dice $S$0, outperforming UNet, SegResNet, VNet, and nnUNet on the 173-patient dataset (Huang et al., 2022). In BIQA, zero-shot SRCC reaches 0.828 on TID2013 and 0.868 on SPAQ, while the reported mean confidence-interval width is reduced relative to LIQE (Lou et al., 25 Jul 2025).

A consistent pattern across domains is that the benefit of DEFNet is largest when fusion is difficult: spatial misalignment in OGMs, low-quality or missing views in multi-view classification, ambiguous pseudo-labels in semi-supervised segmentation, or cross-domain shift in stereo and BIQA. At near-zero conflict or perfect alignment, simple baselines can remain competitive, as explicitly noted for naive DS fusion in occupancy-grid fusion configuration A (Kempen et al., 2023).

7. Limitations, controversies, and open directions

The principal limitation identified across the literature is conflict handling. Classical Dempster fusion can be brittle when sources disagree strongly or are misregistered; this is stated directly for occupancy-grid fusion and motivates learned compensation or conservative alternatives (Kempen et al., 2023). Related papers therefore replace pure normalization by conflict-aware discounting, restricted interaction terms, or contextual reliability. Even so, several methods retain fixed coefficients, such as the $S$1 interaction in IPAF, and the medical segmentation paper explicitly notes that adaptive fusion weights based on local uncertainty or entropy could further improve calibration (He et al., 2024).

Another limitation is that uncertainty quality is unevenly evaluated. Some papers report ECE, Brier, or NLL, especially in multimodal medical segmentation (Huang et al., 2023). Others assess reliability indirectly, such as expert agreement in dual-view remote sensing (Zhao et al., 2023), or do not report formal calibration metrics at all, as stated for EFN and MMEFnet (Zhao et al., 2023, Shi et al., 13 Oct 2025). This suggests that the term "trustworthy" is operationalized differently across DEFNet instantiations.

Scalability is also recurrent. The traffic OGM model is pairwise and single-shot, with no explicit learnable registration module; the paper identifies more agents, temporal fusion, and spatial transformers or correlation layers as logical extensions (Kempen et al., 2023). EDP-MVC uses K-means/class-center imputation and notes that more advanced generative imputation could be integrated (Huang et al., 2024). Co-evidential fusion with IVUM adds per-voxel D-S combination, pignistic transforms, and iterative IVUM computation, which the paper identifies as a source of memory and compute overhead (He et al., 3 Jun 2025). NIG-based systems improve reliability but incur the cost of multi-branch inference, as acknowledged for stereo matching and BIQA (Lou et al., 2023, Lou et al., 25 Jul 2025).

A final conceptual point is that DEFNet is not identical to one evidential formalism. Some instantiations are Dirichlet-EDL systems, some are prototype-based Dempster–Shafer networks, and some are NIG evidential regressors. Taken together, these works suggest a unifying criterion: a model qualifies as DEFNet when deep predictors emit evidential parameters, fusion is performed in evidential space rather than by naive averaging, and uncertainty remains a first-class output rather than a post hoc diagnostic.