Adaptive Multi-Scale Image Processing Using Uncertainty-Aware Deep Feature Fusion

Recent advances in deep learning have significantly improved the performance of image processing systems across a wide range of applications, including image restoration, classification, segmentation, remote sensing, medical imaging, and autonomous systems. These improvements are primarily attributed to the ability of deep learning models to exploit multi-scale representations that capture both fine-grained local structures and high-level semantic information. It has been proven that multi-scale feature learning is especially effective in the cases of complex visual patterns and different object sizes [1], [2], [3].

To enhance feature representation capability, feature fusion mechanisms are widely employed to aggregate information across multiple scales, modalities, and network depths. The classical fusion methods (concatenation, summation, and weighting with attention) have proven to have a promising outcome in various tasks, such as segmentation, classification, and visual retrieval [4], [5], [6]. Nevertheless, the methods tend to make the assumption that all extracted features are equally reliable, which is not true in the real-world context of noise, blur, changes in illumination and low-contrast conditions.

In practical imaging environments, feature reliability may vary significantly across scales due to sensor noise, motion blur, adverse weather conditions, domain shifts, and acquisition artifacts. Traditional multi-scale fusion does not have an explicit process used to quantify and use feature uncertainty, so when unreliable features are prevalent in the fusion process the performance is not optimal or unstable. The application has been noted to be limited in several fields, such as medical imaging [7], [8], remote sensing [9], [10], autonomous driving [11], and image restoration tasks, such as deraining and dehazing [5], [12].

In recent times, uncertainty-aware learning, in which model predictions are provided with estimates of uncertainty confidence, has been emphasized and is being taken seriously. Uncertainty modeling has been effectively used to enhance the robustness, descriptiveness, and generalization in segmentation, detection, and classification tasks [13], [14], [15]. Regardless of these developments, the vast majority of current uncertainty-aware approaches emphasize on the prediction-level calibration or task-specific uncertainty estimation, as opposed to feature-level uncertainty modeling when performing multi-scale fusion. Consequently, unreliable characteristics might continue to have a negative impact on downstream activities.

Several recent studies have explored uncertainty-guided fusion strategies in specific application domains. As an example, uncertainty-aware fusion has been utilized in medical image segmentation [12], [16], [17], face expression recognition [18], fault diagnosis [19] and pansharpening [9]. Similarly, hierarchical and disentangled fusion structures have been proposed to improve robustness and semantic consistency [17], [20].

The research objectives are:

• Develop a Task-flexible Framework: Create a universal, uncertainty-sensitive multi-scale fusion structure that can be seamlessly integrated into various vision tasks—such as image restoration, classification, and segmentation—without requiring significant architectural redesigns.

• Implement Explicit Feature-Level Uncertainty Modeling: Utilize Bayesian deep learning methods, specifically Monte Carlo (MC) dropout, to calculate and represent uncertainty as a “first-class term” at the feature level across different scales.

• Introduce a New Uncertainty-aware Fusion Mechanism: Propose a mathematically-based strategy where multi-scale features are weighted dynamically according to their estimated reliability. This ensures that trusted features have a larger impact on the final representation while less credible, “noisy” information is suppressed.

• Improve Robustness in Adverse Conditions: Enhance the performance, stability, and generalization of image processing systems when operating in adverse imaging conditions involving noise, blur, and low-contrast effects.

Bridge Multi-Scale Learning and Bayesian Models: Fill the existing gap between traditional multi-scale representation learning and Bayesian uncertainty modeling to provide a more interpretable and robust fusion paradigm.

Nevertheless, most existing methods rely on task-specific architectures, specialized supervision strategies, or complex ensemble designs, which limit their general applicability. Furthermore, uncertainty can be implicitly represented by attention or auxiliary losses instead of explicitly being represented by the feature fusion formulation. This inspires the creation of an all-purpose, uncertainty-sensitive multi-scale fusion structure that can be easily incorporated into various vision tasks and which is computationally efficient.

The main aim of this study is to come up with a strong and flexible multi-scale image processing system which clearly considers uncertainty in features during fusion. The main contributions of this paper are summarized as follows:

1. Multi-Scale Framework that is uncertainty-aware: We introduce a single deep learning model that combines image multi-scale decomposition with uncertainty-sensitive learning of features and allows the model to robustly represent the images in adverse imaging conditions.

2. The Uncertainty Modeling at the feature level: The Bayesian deep learning methods, i.e. Monte Carlo dropout are used to calculate uncertainty at a feature level over various scales, a principled metric on feature reliability [21], [22].

3. New Uncertainty-aware Fusion Mechanism: A mathematically-based fusion strategy, in which multi-scale features are weighted dynamically based on the estimated uncertainty, is proposed, such that one can be sure that trusted features have larger contribution to the final representation.

4. Task-flexible Design: The suggested framework can be applied to various image processing problems, such as image restoration, classification, and segmentation, without task architectural redesigns.

5. Comprehensive Evaluation: Large-scale experiments show better performance in comparison to normal fusion and attention-based models on a variety of measures, such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Fréchet Inception Distance (FID), and resilience to synthetic noise and degradation.

At the same time as the current uncertainty-sensitive methods, which base their efforts mainly on prediction calibration or domain adaptation [6], [14], the given method considers uncertainty as a first-class term in multi-scale feature fusion. The representation of the scale-specific features of each feature with a score of uncertainty and the intrinsic conviction of this score in the fusion formulation allows the framework to dynamically adapt to the different levels of image quality and degradation. This architecture bridges the gap between multi-scale representation learning and Bayesian uncertainty model, and offers the general, interpretable and robust fusion paradigm to advance the state of the art in uncertainty-aware image processing.

To clearly differentiate the proposed framework with the current frameworks of uncertainty-aware fusion methods proposed in previous studies [5], [16], several structural and methodological differences can be identified.

Current methods, including the uncertainty-driven deraining model proposed by Tong et al. [5], incorporate uncertainty in multi-scale fusion via task-specific designs, in which uncertainty is implicitly captured in attention modules or auxiliary loss functions. In a similar fashion, the uncertainty-aware adaptive multi-scale U-Net proposed by Sagar et al. [16] introduces uncertainty estimation at the decoder level in the segmentation task, based on hierarchical skip connections and task-specific supervision.

Instead, the suggested approach presents the feature-level uncertainty as a well-defined and modular constituent of the fusion process that is not dependent on task-specific architectures. The main differences are as follows:

• Level of Integration: Earlier models focus uncertainty at the prediction or decoder level, but the framework proposed in this paper directly models uncertainty at the feature level, at all scales.

• Fusion Mechanism: The current approaches are based on attention or learned weights, whereas the suggested approach is based on a closed-form uncertainty-weighted fusion formulation, which allows interpretable and probabilistically based feature aggregation.

• Task Generality: The proposed framework is task-flexible, it can be easily incorporated in restoration, classification, and segmentation pipelines and does not require a redesign of the architecture, unlike the task-specific architectures proposed in previous uncertainty-aware fusion studies [5], [16].

• Mathematical Formulation: The given approach explicitly uses the uncertainty as part of the fusion weights through inverse-variance relationship, which offers a principled Bayesian interpretation, as opposed to heuristic or learned attention mechanisms.

This mathematical and structural difference confirms the originality of the suggested uncertainty-aware fusion strategy as a generalized, interpretable, and theoretically-based alternative to currently used techniques.

To provide a clearer overview of existing research, the literature is categorized into three primary directions: multi-scale representation learning, uncertainty modeling, and feature fusion strategies.

Multi-scale feature learning has been extensively investigated as an effective approach for capturing both local structural details and global semantic information in images. Early deep learning models primarily relied on fixed-scale representations and therefore struggled to handle varying object scales and complex visual patterns. To overcome this, multi-scale decomposition and aggregation models have been popularly used in remote sensing, medical imaging and scene understanding tasks. As an example, Cai et al. [1] showed that the performance of fine-scale grassland classification can be enhanced significantly by combining multi-sensor images obtained by UAVs using deep multi-scale representations. In the same vein, Ding et al. [2] also proposed Multi-scale Perception and Context Network (MPCNet), a model to combine multi-scale decomposition and context-gating fusion to achieve stronger robustness in medical image classification.

In other contexts, hierarchical multi-scale feature aggregation has been developed to enhance semantic consistency and accuracy of retrieval. Cui et al. [4] presented a hierarchical semantic aggregation framework that utilizes multi-scale representations to improve visual retrieval performance through organized feature fusion. Another domain-aware multi-scale fusion architecture, MulFF-Net, was proposed by Halder and Mahadevappa [3] for medical image segmentation, where the authors demonstrated the effectiveness of scale-adaptive fusion in the medical image segmentation field, and the authors discovered that scale-adaptive fusion is effective.

Uncertainty-aware learning has become an essential paradigm for improving the robustness, reliability, and interpretability of deep vision models. By modeling both epistemic and aleatoric uncertainty, deep networks can estimate prediction confidence, particularly in noisy or ambiguous environments. According to Senousy et al. [7], Multi-level Context and Uncertainty-aware dynamic deep ensemble (MCUa) is a dynamic deep ensemble, which combines uncertainty estimation with multi-level contextual learning, to classify histology of breast cancer, with higher robustness and calibration.

Uncertainty modeling has been integrated into multiple vision problems within the past years, such as semantic segmentation, object detection, and domain adaptation. Yin et al. [14] proposed an uncertainty-aware domain adaptive semantic segmentation framework capable of mitigating performance degradation caused by domain shifts. Similarly, Zhang et al. [22] introduced a scale-disentangled and uncertainty-guided alignment strategy for domain-adaptive object detection and demonstrated the usefulness of uncertainty cues in cross-domain environments coined scale-disentangled and uncertainty-guided alignment to object detection in the domain of domain adaptation, and showed the usefulness of uncertainty cues in cross-domain environment.

Uncertainty estimation has been considered as an integral part of multi-scale fusion and has been receiving growing interest, especially in medical imaging and image restoration problems. Lyu et al. [23] suggested an adaptive feature aggregation framework for uncertainty-guided semi-supervised medical image segmentation, where uncertainty estimates are integrated into multi-task learning. Based on this concept, Sagar et al. [16] created an uncertainty-aware adaptive multi-scale U-Net for low-contrast cardiac image segmentation an uncertainty-aware adaptive multi-scale U-Net to segment low-contrast cardiac images, which demonstrated a high level of performance in the imaging of complex scenarios.

A few studies have examined uncertainty-based fusion-based segmentation and restoration. Dong et al. [13] proposed a Unified Uncertainty-Aware Feature Reassembly model (UAFer) for class-agnostic binary segmentation. Wang et al. [24] introduced Dual Perception Guided Network (DPGNet), which incorporates uncertainty perception to improve boundary segmentation. Similarly, Zhang et al. [15] introduced the Uncertainty-Aware Network (UANeT), which performs feature calibration and boundary refinement during medical image segmentation.

In image restoration, Tong et al. [5] suggested an uncertainty-based multi-scale fusion network to restore images using deraining and Wang et al. [8] proposed an uncertainty-based weakly-supervised-dehazing network which incorporates dynamic attention and multi-scale feature fusion. These experiments show that the fusion that is uncertainty-aware can substantially enhance the quality of restoration in unfavorable visual-conditions.

Multi-scale fusion with uncertainty awareness has also been applied in other areas of application successfully with various applications other than medical imaging. Li and Singh [11] suggested a multi-modal fusion framework, in which they utilized uncertainty-aware learning, to enhance the obstacle detection in adverse weather conditions in autonomous driving. Lian and Li [19] applied uncertainty-aware multi-scale fusion in fault detection of load imbalance in isolators of the railway tracks in fault diagnosis. In detection of landslides, Wang et al. [10] examined the impact of uncertainty modeling in deep networks incorporating multi-scale information, and it is worth noting that uncertainty modeling leads to high detection reliability.

Uncertainty-guided fusion strategies have also been useful with remote sensing and geospatial analysis. Zheng et al. [9] suggested a deep adaptive pansharpening structure that combines the uncertainty-aware image fusion to improve the spatial-spectral quality. Yang et al. [25] provided a method of building change detection based on the fusion of multi-scale features with the awareness of uncertainty, which enhances the accuracy of detection in complicated urban settings.

Recently, more advanced uncertainty-aware fusion paradigms, including disentangled, hierarchical, and federated learning-based frameworks, have been investigated. DFuse-Net uses disentangled feature fusion, which is combined with uncertainty-aware learning to ensure effective multi-modal brain tumor segmentation [17]. The hierarchical aggregation network uses uncertainty modeling as the means to enhance the segmentation performance [20]. Zhao et al. [26] proposed a multi-modal cell image segmentation framework where the uncertainty in each of the modalities is a learned aspect of the framework under consideration.

In addition to centralized learning, Xiong et al. [21] suggested AdaptiveStreamFL, a multi-scale federated learning framework enhanced with uncertainty quantification based on the Bayesian methodology, and proved the uncertainty-aware multi-scale learning scalability to dynamic data settings. Concerning image classification, Zhang et al. [6] examined uncertainty-aware decision fusion to strengthen adaptive deep networks to support the generality of the uncertainty-guided fusion strategies.

Recent studies published in the International Journal of Computational Methods and Experimental Measurements have also highlighted the growing importance of intelligent image processing and computational modeling techniques in complex engineering applications. Haddadi [27] investigated adaptive quality-energy trade-offs in image processing using statistical priority classification and variable-approximate computing, demonstrating the effectiveness of adaptive computational frameworks in improving image quality and processing efficiency. In addition, recent computational measurement studies in the journal have emphasized the significance of robust data-driven modeling and intelligent optimization strategies for improving computational accuracy and system reliability. These contributions further support the relevance of integrating uncertainty-aware and adaptive feature fusion techniques into modern image processing systems.

Moreover, recent studies published in high-impact journals have emphasized the importance of uncertainty-aware learning in complex vision tasks. Works in journals such as IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Image Processing, and Medical Image Analysis have demonstrated the effectiveness of probabilistic modeling and uncertainty-guided learning in improving robustness and generalization. These studies further motivate the need for integrating uncertainty directly into feature fusion mechanisms.

Even though significant advances have been achieved in the multi-scale feature learning and uncertainty-sensitive modeling, the current solutions are either task-oriented, domain-specific or based on complex architectural designs. In the methods many will use uncertainty at the prediction level or the decision level, instead of explicitly using that in the formulation of feature fusion. In addition, current uncertainty-sensitive fusion methods are frequently task-specific (in terms of restoration, classification and segmentation).

Conversely, the current paper will attempt to address this gap by providing a task-flexible, uncertainty-aware multi-scale feature fusion model that explicitly represents the uncertainty at the feature level and incorporates it into a principled feature fusion process. The proposed approach by dynamically weighting multi-scale features based on their estimated uncertainty makes the state of the art in robust and adaptive image processing.

This section presents the proposed Adaptive Multi-Scale Image Processing framework using Uncertainty-Aware Deep Feature Fusion. The proposed framework is designed to be task-flexible and robust against adverse imaging conditions by explicitly modeling uncertainty during multi-scale feature fusion. The conceptual outline of the structure is presented in the form of the pipeline that comprises multi-scale image decomposition, deep feature extraction, uncertainty estimation, uncertainty-aware fusion, and task-specific prediction.

Given an input image $I \in R^{H \times W \times C}$, the framework first decomposes the image into multiple scale representations $\left\{I_l\right.$, $\left.I_2, \ldots, I_N\right\}$, where each scale captures complementary spatial and frequency information. The discriminative features are obtained at each scale by a deep encoder and the uncertainty estimated by a Bayesian inference mechanism. These scalespecific features are then combined using a new uncertaintyaware weighting strategy to come up with a strong global representation which is finally inputted to a task-specific head to restore images, classify or segment them.

In order to give a direct and easily comprehensible explanation about the proposed uncertainty-aware multi-scale framework, Figure 1 depicts the general architecture of the system. The model is made up of consecutive steps involving multi-scale image de-composition, heavy feature extraction at every scale, Bayesian uncertainty estimation, uncertainty-concerned feature fusion, and task-specific prediction heads. Information on uncertainty is clearly included in this flowchart as it is part of the fusion process to increase the robustness and reliability of the system.

Instead, in order to obtain structure of images in the various resolutions and frequencies, the input image is diluted into numerous scales that are called Laplacian Pyramid decomposition or Wavelet Transform-based decomposition.

Laplacian Pyramid decomposes an image into a series of band-pass filtered images and low-frequency residual image. The Gaussian pyramid is built on paper as follows formually:

where, $i=0,1, \ldots, \mathrm{n}-1$. Laplacian representation at scale iii is then obtained as:

where, $L_i$ captures fine-to-coarse details one after another thus letting the network to extract edges, textures, and global structures individually.

In place of this, discrete wavelet transform (DWT) or stationary wavelet transform (SWT) is used to divide the input image into sub-bands relating to various frequency orientations:

where, $LL$ denotes low-frequency components. $LH$, $HL$, and $HH$ represent high-frequency details. Wavelet-based decomposition is also known to be useful in capturing directional textures and preserving spatial information despite noise and blur.

The process of multi-scale image decomposition and the following process of deep feature extraction at each scale is visualized in Figure 2. The framework identifies complementary spatial details by breaking down the input image into several frequency levels or resolution levels. Deep encoders individually work on each scale, which allows learners to effectively learn representations in the face of a wide range of image degradations.

To each of the decomposed scales $I_i$, one uses a deep feature extractor $E_i(\cdot)$ to get scale specific representations:

Two other encoder designs are taken into consideration:

• Lightweight Convolutional Neural Network (CNN) encoders: They are shallow convolutional blocks that have residual connection that help in extracting local spatial features at low costs of computation.

• Encoders: Vision Transformer (ViT) encoders: ViT-based encoders are used to capture long-range dependencies, in which each scale image is divided into patches and passed through self-attention layers.

The encoders may exchange parameters both between scales to scale-down the model or use scale-dependent weights to increase representational diversity.

To estimate feature-level uncertainty, MC dropout is employed as an approximation to Bayesian inference. In the proposed framework, dropout layers are inserted after convolutional or transformer encoder blocks at each scale, with a dropout rate typically set between 0.2 and 0.5.

During inference, dropout remains active, and $T$ stochastic forward passes are performed for each input (empirically set to $T$ = 5–10). For each scale iii, this produces a set of feature realizations:

Then For each scale $i$, $T$ stochastic predictions are then obtained:

The average representation of the feature is calculated as:

where, the uncertainty that it puts about is the uncertainty that is daydreamed:

The predictive mean is the average feature embedding, and the variance is the epistemic uncertainty due to uncertainty in the model parameters.

Computational cost:

MC dropout brings in a linear cost of inference based on $T$. The overall overhead is however manageable, owing to lightweight encoders and common weights across scales. Practically, $T$ = 5 is a good trade-off between the quality of uncertainty estimation and computation efficiency.

This design guarantees that the estimation of uncertainty is not only scalable but also can be applied to the real world.

The basic contribution of this work is an uncertainty-aware fusion mechanism, which relies on the principled dynamically-weighted fusion of multi-scale features, based on their uncertainty estimation.

Fusion weights are determined as given the mean feature vectors, i.e. $\left\{\mu_i\right\}_{\{i=1\}}^{\{N\}}$ and uncertainty estimates, i.e. $\left\{\sigma_i\right\}_{\{i=1\}}^{\{N\}}$.

where, $\varepsilon$ is a small constant to ensure that it is numerically stable.

Eq. (9) is based on the concepts of the inverse-variance weighting in Bayesian estimation, in which observations with less uncertainty (variance) are value-weighted to greater importance. It is frequently applied in probabilistic sensor fusion and statistical estimation, whereby a combination of many estimates is found to provide the best result by inverting the uncertainty in each estimate.

Here, the values of the uncertainties are The $\sigma_i^2$ are squared, and serve as measures of reliability of scale-specific properties. This formulation will make the fusion mechanism theoretically-grounded, so that the more confident features play a more important role in the final representation. The formulas are similar and uncertainty-aware learning and probabilistic fusion have been investigated [9], [21].

$F$ final fused representation is calculated as:

This formulation makes sure that more weight is assigned in the final form (greater uncertainty) to high confidence features whereas unreliable features are automatically down-weighted. In contrast to attention-based fusion, it offers a fusion strategy that is probabilistically based and is interpretable.

In order to highlight the essence of the contribution of the work, Figure 3 demonstrates the uncertainty-aware fusion mechanism of features. An uncertainty estimate of each of the scale-specific features is obtained by a Bayesian inference. The uncertainty values are computed dynamically, by which the fusion weights are calculated, so that the reliable features will have a greater contribution to the final representation.

The proposed fusion formulation can be interpreted from a probabilistic perspective. Specifically, the inverse-uncertainty weighting in Eq. (9) is analogous to minimum-variance estimation, where multiple noisy observations are combined to produce an optimal estimate.

Assuming that each scale-specific feature represents an independent estimate with associated uncertainty, the fusion process minimizes the overall variance of the combined representation. This leads to a solution where weights are proportional to the inverse of the variance, ensuring that more reliable features contribute more significantly.

Computationally, the fusion step is linear with $N$, the number of scales, and the uncertainty estimation is multiplicative with a factor $T$, as a result of stochastic sampling. Thus, the complexity of the general form can be stated as: O($N$ × $T$ × $F$), where $F$ is cost of feature extraction.

This formulation demonstrates that the proposed method is both computationally tractable and theoretically grounded, making it suitable for practical deployment.

This fused feature representation $F$ is then inputted into task-specific heads, which allows the framework to be used on a variety of tasks of image processing:

• Image Restoration: $F$ is used to construct a clean image by a decoder network, which is optimized on pixel-wise and perceptual loss functions.

• Image Classification: The next layer consists of fully connected layers then a softmax activation to generate class probabilities.

• Image Segmentation: Upsampling and skip connection decoder produces fine pixel predictions.

This design is compatible with modularity in the sense that uncertainty-aware fusion backbone can be reused in a task with relatively few architectural reconfigurations.

It is important to note that the proposed framework follows a shared backbone with task-specific heads. The multi-scale feature extraction, uncertainty modeling, and fusion modules are fully shared across tasks, while only the final prediction layers are task-dependent.

No task-specific tuning is applied to the fusion mechanism, ensuring that the learned uncertainty-aware representation remains generalizable across different vision tasks. This architecture emphasizes the flexibility of the suggested method and its flexibility in the tasks without the need to retrain or change the structure of every task.

Although the framework is meant to be universal in various tasks, the evaluation in hand targets representative tasks in restoration, classification and segmentation. More validation in other fields is a significant line of future research.

The same backbone configuration, including the multi-scale feature extractor and uncertainty-aware fusion module, is consistently used across all tasks, with only minor adjustments in the final prediction heads to accommodate task-specific output requirements.

The end-to-end training of the framework is done with a composite loss:

where, $L_{\text {task}}$ is task-dependent (e.g., cross-entropy, Dice loss or $\ell 1 / \ell 2$ loss), and $L_{\text {reg}}$ may act to regularize the uncertainty estimates to avoid degenerate solutions. Hyperparameter $\lambda$ determines the influence of regularization term.

The uncertainty estimation forward pass computational overhead is mainly caused by both the $T$ stochastic forward passes in training and inference. In practice a small value (e.g., $T$ = 5–10) is an excellent compromise between performance and efficiency. The framework is also appropriate to use in practice because of lightweight encoders and shared weight of scales that decrease the computational cost.

The proposed framework is trained using standard stochastic gradient-based optimization, and convergence behavior is similar to conventional deep learning models with comparable architectures.

Convergence Behavior:

Empirically, the model converges within 80–100 epochs across all tasks. The inclusion of uncertainty estimation does not significantly affect convergence stability, as MC dropout is only activated during inference and does not alter the deterministic training process.

Scalability:

The computational complexity scales linearly with:

• Number of scales $N$;

• Number of stochastic passes $T$.

This ensures that the model can be adapted to different computational budgets by adjusting these parameters.

Efficiency Considerations:

• Shared encoder weights reduce memory footprint;

• Lightweight architectures enable faster training;

• Adjustable stochastic sampling allows dynamic trade-offs.

These properties make the framework suitable for both high-performance and resource-constrained environments.

Overall, the suggested approach combines the multi-scale image decomposition, deep feature mining, Bayesian uncertainty model and a new uncertainty-based fusion system into a single system. The approach is robust and more interpretable and exhibits generalization in a variety of image processing tasks and adverse image perception conditions by expressly considering the reliability of features during fusion.

This part of the paper gives a detailed analysis of the proposed adaptive multi-scale image processing using uncertainty-aware deep feature fusion framework. The analysis is designed to emphasize on quantitative performance, ability to withstand degradation, ablation studies, and qualitative information.

The proposed framework is evaluated on multiple benchmark datasets across different tasks:

• Image Restoration: BSD68, Set12;

• Image Classification: CIFAR-10, CIFAR-100;

• Image Segmentation: ISIC 2018, Cityscapes.

Training details:

• Optimizer: Adam;

• Learning Rate: 1 × 10$^{\text{-4}}$ (with cosine decay);

• Batch Size: 16;

• Number of Epochs: 100;

• Dropout Rate: 0.3;

• Number of MC Samples: $T$ = 5.

Loss functions:

• Restoration: $\ell 1$ + perceptual loss;

• Classification: Cross-entropy;

• Segmentation: Dice + cross-entropy.

Hardware:

Experiments were conducted on NVIDIA GPUs (e.g., RTX 3090) using PyTorch.

All models were trained under identical conditions to ensure fair comparison.

Baselines:

To ensure a fair and reproducible comparison, the proposed method is evaluated against well-established and representative models from different fusion paradigms:

• Single-Scale CNN: ResNet-18-based encoder (no multi-scale fusion);

• Multi-Scale Concatenation: Feature Pyramid Network (FPN);

• Attention-Based Fusion: CBAM (Convolutional Block Attention Module) and Attention U-Net;

• Uncertainty-Aware (No Fusion): MC dropout-based single-scale model;

• Task-Specific State-of-the-Art:

(a) Image Restoration: U-Net, DnCNN;

(b) Image Segmentation: U-Net++, DeepLabV3+;

(c) Image Classification: ResNet-50, EfficientNet-B0.

The choice is made so that the comparison is related to classical, attention-based and uncertainty-aware paradigms and this way offers a detailed assessment of the suggested method.

Dataset Scale and Splits:

To have a thorough assessment, datasets of different sizes and complexities are utilized. Specifically:

• BSD68: 68 test images (typically used as denoising benchmarks);

• Set12: 12 standard grayscale images for restoration evaluation;

• CIFAR-10: 60,000 images (50,000 training/10,000 testing);

• CIFAR-100: 60,000 images across 100 classes;

• ISIC 2018: ~10,000 dermoscopic images for segmentation;

• Cityscapes: 5,000 finely annotated urban scene images.

These datasets offer a combination of small-scale testing datasets and large-scale real-world datasets, such that the test is both controlled and representative.

All reported results are averaged over three independent runs with different random initializations. To ensure fairness, consistent training settings are maintained across runs, while random seeds are varied to capture performance variability.

Table 1 presents the comparison of the performance of the proposed method of restoring the images with fusion strategies on the basis line in different noise conditions.

All reported results are averaged over three independent runs, with standard deviation included to reflect performance stability.

The suggested approach has the largest PSNR and SSIM values, thus having excellent reconstruction quality. The contrast to the attention-based fusion shows that it is more effective to introduce uncertainty into a fusion mechanism instead of using only learned attention weights.

To provide a clearer visualization of performance variability, Figure 4 includes error bars representing standard deviation across multiple runs.

To compare between perceptual quality, FID scores were also calculated between restored and ground-truth images. The perceptual image quality comparison results are summarized in Table 2.

The reduced FID score of the proposed framework shows it to be less damaged to image statistics and perceptual realism, especially in difficult cases of degradation.

One of the primary objectives of this work is to maintain robustness under uncertainty-inducing conditions. Table 3 indicates a decrease in relative performance with increases in noise and blur severity.

The uncertainty-aware fusion model exhibits significantly lower performance degradation, demonstrating the effectiveness of suppressing unreliable features during the fusion process.

Figure 5 shows the tendencies of strength of the various fusion strategies with the rising noise levels. The proposed method demonstrates a substantially slower degradation rate, confirming its ability to suppress unreliable features through uncertainty-aware weighting.

In addition to performance accuracy, the computational efficiency of the proposed framework is evaluated in terms of inference time, model size, and scalability.

Due to the use of MC dropout, the proposed method requires multiple stochastic forward passes. Table 4 compares the average inference time per image:

As expected, the runtime increases approximately linearly with the number of stochastic passes $T$. However, using $T$ = 5, which provides near-optimal performance, maintains acceptable inference speed for practical applications. The comparison of model sizes and parameter counts is presented in Table 5.

The proposed framework maintains a relatively compact model size due to lightweight encoders and shared weights across scales.

Although the proposed method introduces additional computational overhead due to uncertainty estimation, this cost is justified by the significant gains in robustness and performance under degraded conditions. Moreover, the framework allows flexible adjustment of $T$, enabling a trade-off between efficiency and accuracy.

An ablation study was conducted to evaluate the contribution of each framework component. The ablation study results evaluating the contribution of each framework component are summarized in Table 6.

The results show that:

• Multi-scale decomposition provides a performance boost, obviously.

• More robustness is added by uncertainty modeling.

• The greatest advantage is realized in the scenario of uncertainty that is directly introduced into the fusion mechanism, which proves the nature of the present work.

The comparison in Table 7 isolates the effect of the fusion mechanism by using the same backbone architecture. The results demonstrate that uncertainty-aware fusion provides more reliable feature weighting than attention-based fusion, particularly in the presence of noise and degraded inputs.

In order to further test the strength of the suggested framework, key hyperparameters such as the number of stochastic forward passes ($T$) and the number of scales of decomposition ($N$) were sensitivity analysed. The sensitivity analysis with varying MC dropout samples is presented in Table 8.

Results indicate that performance increases as $T$ increases but flattens at $T$ = 5, reflecting a good trade-off between the cost of computation and accuracy. The effect of the number of decomposition scales on performance is summarized in Table 9.

There is a slight improvement in using more than three scales but the complexity of computation increases, indicating that $N$ = 3 is a good option.

This discussion shows that the given framework is robust in varying parameter conditions, which supports its practical use.

The outcomes of the experiments repeatedly prove that the suggested framework is more effective than the traditional multi-scale and attention-based fusion strategies in all the considered tasks and measures. In contrast to the attention mechanisms that implicitly learn the importance weights, the proposed approach proposes a probabilistically-based weighting scheme that is motivated by the uncertainty at the feature level. It results in a higher interpretability and stability especially in unfavorable imaging scenarios.

The task-independent nature of the framework is also another advantage worth pointing out. This uncertainty-aware backbone of fusion was also used with high success in the restoration, classification as well as segmentation without structural changes, which is an indication of the high ability to generalize.

Practically, the computational penalty of Monte Carlo dropout is small and warrants the great improvements in strength and performance. Besides, the uncertainty measures that the framework produces contain important information about model confidence that is essential to safety-sensitive systems like medical imaging and autonomous systems.

Upon further examination, it can be seen that the changes in performance are not consistent in all conditions but are significant in high uncertainty situations. In particular, the most significant improvements in the proposed approach are on:

• High noise levels: where unreliable fine-scale features are successfully suppressed.

• Motion blur: in which the consistency of features is degraded by scale.

• Low-contrast conditions: in which uncertainty estimation is useful in determining informative regions.

The performance improvement is moderate in comparatively clean settings, which means that the main benefit of the proposed framework is the possibility to adjust to the degraded or ambiguous inputs.

Moreover, the uncertainty-guided weighting offers more predictable and consistent performance as compared to attention-based fusion, which learns weights implicitly, especially when the input distribution is not consistent with training conditions.

This implies that uncertainty-sensitive fusion will be particularly useful in practical applications where noise and shift to domain cannot be avoided.

For example, in the Gaussian noise experiments ($\sigma$ = 25), fine-scale features tend to become unreliable due to noise amplification. In such cases, the proposed method assigns lower weights to high-uncertainty fine-scale representations and relies more on stable coarse-scale features. This adaptive behavior leads to improved reconstruction quality, as reflected in the higher PSNR values compared to attention-based fusion methods.

To conclude, the suggested uncertainty-sensitive multi-scale fusion system:

• High-performance in state-of-the-art in several different measures.

• Extremely strong in the presence of noise, blur and low-contrast.

• Fusion of interpretable features based on uncertainty.

• Generates effective generalizations to a variety of image processing problems.

These results confirm that explicit uncertainty modelling, which is realized on the feature fusion level, is a powerful and generic way of robust deep image processing.

This article introduced a multi-scale feature fusion system that is task flexible and uncertainty-aware to perform robust image processing. The proposed approach directly incorporates Bayesian uncertainty estimation into the feature fusion process, which serves as an effective way to overcome the shortcomings of standard fusion strategies, which assume that features are as reliable as each other.

The experimental findings on various tasks show that the framework always enhances the performance, especially in demanding environments like noise, blur and low contrast. The probabilistic version of the fusion mechanism is also more interpretable and stable than the attention-based methods.

Computationally, the approach creates moderate overheads associated with stochastic sampling, yet is efficient because of its lightweight design and choice of parameters that can be scaled.

Although it has its benefits, the proposed framework has a number of limitations. First, MC dropout makes the inference time more dependent on the MC dropout, which could hinder real-time usage. Second, the analysis is only done on a collection of representative datasets and further validation on larger scopes would enhance the generality of the method. Lastly, the independent uncertainty across scales assumption might not be sufficient to account for complex feature correlations.

Future research will aim at minimizing the computational burden, other methods of estimating uncertainty, and generalizing the framework to other applications in the real world.

These findings highlight the potential of uncertainty-aware multi-scale feature fusion as a generalized and interpretable framework for next-generation intelligent image processing systems.