TaskIR: Task-Driven Image Restoration via Degradation Adaptation and Task Feedback

1 INTRODUCTION

• Shared restoration transformations may not adapt sufficiently to distinct degradation demands, and better image quality alone does not guarantee better downstream task performance.Heterogeneous task features can also interfere with restoration, while indiscriminate feedback refinement may disrupt already restored content.Source: §1 · p. 2
• TaskIR combines degradation-adaptive restoration and task feedback refinement in a two-stage framework for diverse degradation conditions.The framework aims to balance restoration quality and downstream task performance.Source: §1 · p. 2
• DGTB uses degradation representations to dynamically modulate feature transformations in a shared restoration network for adaptive restoration across degradations.The framework's DRM jointly models global degradation characteristics and spatial degradation responses.Source: §1 · p. 2
• TRFG transforms heterogeneous task features into targeted feedback, while STFR selectively refines restoration features to limit interference with restored content.The feedback is generated based on discrepancies in task representations of current restoration results.Source: §1 · p. 2
• Experiments across diverse degradations and downstream vision tasks report improved task performance while maintaining high restoration quality.The paper presents this as evidence that TaskIR balances both objectives under varied degradation conditions.Source: §1 · p. 2

2 RELATED WORK

• All-in-one restoration methods address multiple degradations within one model, using approaches such as degradation representations, prompts, and dynamic experts.Examples include AirNet, PromptIR, and MoCE-IR; later work explores feature decomposition, bio-inspired visual modeling, and degradation-content decoupling.Source: §2 · p. 2
• Existing all-in-one methods mainly optimize pixel fidelity and visual quality, with limited attention to whether restored images suit downstream tasks.Source: §2 · p. 3
• Task-driven restoration methods use downstream constraints for applications including super-resolution, dehazing, and underwater enhancement, but most target a single degradation.Task representations may not align with restoration requirements, and indiscriminate task-guided refinement may introduce irrelevant information and interfere with restored content.Source: §2 · p. 3

3.1 STAGE I: DEGRADATION-GUIDED RESTORATION

• The DRM extracts global degradation characteristics and spatial responses, which the DGTB uses to modulate shared-network feature transformations across restoration stages.The DRM derives its global representation from pooled mean and dispersion statistics; the DGTB also adjusts update magnitudes according to spatial degradation maps.Source: §3.1 · p. 4

3.2 STAGE II: TASK-DRIVEN FEEDBACK MODELING

• Stage II uses task feedback to refine Stage I restoration, with TRFG generating feedback and STFR selectively updating intermediate restoration features.STFR reduces interference with already restored content through feedback fusion and selective feature refinement.Source: §3 · pp. 3–4, §3.2 · pp. 4–5

3.2.1 TASK-TO-RESTORATION FEEDBACK GENERATION

• TRFG aligns heterogeneous task features, adapts them with task-conditioned multi-scale branches, and predicts the current restoration’s task-representation discrepancy.During training, an EMA teacher provides discrepancy targets from the initial restored and reference images; inference does not require the reference image.Source: §3.2.1 · p. 5

3.2.2 SELECTIVE TASK FEEDBACK REFINEMENT

• STFR weights feedback signals by their relevance to the current restoration and uses conditional refinement to control feature updates.It compares candidate update and keep states under the same feedback to adaptively control refinement and reduce interference with already restored content.Source: §3.2.2 · pp. 6–7

4.1 IMPLEMENTATION DETAILS

• Stage I and Stage II are trained for 200K and 100K iterations, respectively; in Stage II, only TRFG and STFR are optimized while the restoration backbone and task networks are frozen.Stage I uses pixel-wise L1 reconstruction and degradation-classification losses.Source: §4.1 · p. 7

4.2 COMPARISONS WITH THE STATE-OF-THE-ART METHODS

• TaskIR achieves competitive PSNR and SSIM across eight degradation types while retaining high restoration quality after task feedback is added.Qualitative comparisons describe sharper recovered boundaries and structural details than competing methods.Source: §4.2 · p. 8, Figure 3 · p. 8
• TaskIR achieves competitive performance across classification, segmentation, and detection metrics, with qualitative results showing better-aligned predictions and preserved object boundaries.The reported metrics include Top-1/Top-5 accuracy, mIoU/Dice, and mAP.Source: §4.2 · pp. 8–9, Figure 4 · p. 8

4.3 ABLATION STUDY

• Adding the spatial degradation map improves restoration quality over using only the global degradation representation, while task feedback improves task performance at comparable restoration quality.The ablation compares C1 (global representation with feedback), C2 (full TaskIR), and C3 (without feedback).Source: §4.3 · p. 9
• The full feedback-module configuration performs better across downstream task metrics than configurations missing an individual module, while maintaining comparable restoration quality.The ablations indicate that feedback fusion and selective feature refinement both contribute to task-feedback utilization.Source: §4.3 · p. 9

• TaskIR jointly models degradation adaptation and task feedback through a two-stage image restoration design.Source: §5 · p. 10
• Degradation representations modulate Transformer feature transformations to adapt restoration across diverse degradations.Source: §5 · p. 10
• Task feedback selectively refines restoration features to improve downstream suitability while reducing interference with already restored content.Source: §5 · p. 10