PaperlayerView exact revision on arXiv

Source-linked AI summary

Pruning Overparameterized Multi-Task Networks for Degraded Web Image Restoration

Thomas Katraouras, Dimitrios Rafailidis

arXiv:2510.14463v1cs.CV

TL;DR

Online social networks degrade web images, while multi-task restoration models can handle diverse degradations but often require many parameters. MIR-L uses iterative lottery-ticket pruning, removing low-magnitude weights and resetting survivors to their original initialization. On restoration benchmarks, it reduces trainable parameters by up to 90% while maintaining high performance.

  • Problem

    Multi-task restoration models handle diverse image degradations but often require many parameters, and the effectiveness of lottery tickets for compressing them has not been thoroughly explored.

    Source: §I · pp. 1–2

  • Method

    MIR-L iteratively prunes low-magnitude weights from multi-task restoration models and resets surviving weights to their original initialization.

    Source: §I · p. 2

  • Results

    MIR-L reduces trainable parameters by up to 90% while maintaining high benchmark performance across deraining, dehazing, and denoising tasks.

    Source: §I · p. 2

  • Takeaways & Limitations

    The sparse subnetworks often match or exceed dense state-of-the-art multi-task restoration models while using fewer trainable parameters.

    Source: §I · p. 2

Abstract

from arXiv · show

Image quality is a critical factor in delivering visually appealing content on web platforms. However, images often suffer from degradation due to lossy operations applied by online social networks (OSNs), negatively affecting user experience. Image restoration is the process of recovering a clean high-quality image from a given degraded input. Recently, multi-task (all-in-one) image restoration models have gained significant attention, due to their ability to simultaneously handle different types of image degradations. However, these models often come with an excessively high number of trainable parameters, making them computationally inefficient. In this paper, we propose a strategy for compressing multi-task image restoration models. We aim to discover highly sparse subnetworks within overparameterized deep models that can match or even surpass the performance of their dense counterparts. The proposed model, namely MIR-L, utilizes an iterative pruning strategy that removes low-magnitude weights across multiple rounds, while resetting the remaining weights to their original initialization. This iterative process is important for the multi-task image restoration model's optimization, effectively uncovering "winning tickets" that maintain or exceed state-of-the-art performance at high sparsity levels. Experimental evaluation on benchmark datasets for the deraining, dehazing, and denoising tasks shows that MIR-L retains only 10% of the trainable parameters while maintaining high image restoration performance. Our code, datasets and pre-trained models are made publicly available at https://github.com/Thomkat/MIR-L.

I. INTRODUCTION

Online social networks degrade images through lossy operations, while multi-task restoration models address varied degradations but remain parameter-heavy. MIR-L applies lottery-ticket pruning to these models; experiments report up to 90% fewer trainable parameters while maintaining high restoration performance.

I. INTRODUCTION

  • Multi-task restoration models handle multiple degradation types in one framework without prior knowledge of the degradation, but often require many parameters.Their parameter demands raise computational and memory costs, limiting use on consumer devices and in real-time applications such as client-side web restoration.Source: §I · pp. 1–2
  • The paper investigates Lottery Ticket Hypothesis subnetworks for multi-task restoration, a setting whose effectiveness had not been thoroughly explored.The hypothesis describes small subnetworks in overparameterized networks that can train in isolation to achieve performance comparable to the original model.Source: §I · p. 2
  • MIR-L iteratively removes low-magnitude weights and resets surviving weights to their original initialization as part of multi-task model optimization.The method targets deraining, dehazing, and denoising.Source: §I · p. 2
  • Global pruning finds very sparse winning tickets, whereas layer-wise pruning diminishes performance.The paper explores both strategies for multi-task restoration models.Source: §I · p. 2
  • MIR-L reduces trainable parameters by up to 90% while maintaining high performance across benchmark deraining, dehazing, and denoising tasks.The sparse networks often match or exceed dense state-of-the-art multi-task models and outperform baseline pruning methods.Source: §I · p. 2

II. PRELIMINARIES

The preliminaries define pruning and distinguish magnitude-based, one-shot, and iterative approaches by how they select and remove network parameters.

II. PRELIMINARIES

  • Pruning removes selected network connections to reduce parameter count and seek lower memory and computational costs while preserving performance.Formally, it removes a subset of parameters from a dense network to produce a sparse network.Source: §II · p. 2

1) Magnitude Pruning.:

  • Magnitude pruning removes parameters whose absolute values fall below a threshold, assuming they contribute less to performance.Pruned parameters are set to zero.Source: 1) Magnitude Pruning · p. 2

2) One-shot Pruning.:

  • One-shot pruning removes a fixed percentage of parameters once, after the initial training phase.The pruning criterion can be magnitude-based, and removed parameters are set to zero.Source: 2) One-shot Pruning · p. 2

3) Iterative Pruning:

  • Iterative pruning sparsifies a network over multiple rounds, optimizing it after each pruning step.Each round prunes a percentage of parameters.Source: 3) Iterative Pruning · p. 2

III. PROPOSED METHOD

The proposed model addresses deraining, dehazing, and denoising without prior knowledge of the degradation. Its UNet-style encoder-decoder uses transformer blocks and prompt modules to incorporate degradation-aware information during reconstruction.

1) Tasks:

  • The blind multi-task model is designed to recover clean images from degraded inputs without prior knowledge of the degradation type.It handles deraining, dehazing, and denoising.Source: 1) Tasks · p. 2

2) Architecture:

  • A UNet-style encoder-decoder with transformer blocks extracts and refines features, while sequential prompt blocks inject degradation-aware information during decoding.The model takes a degraded RGB image as input; prompt generation weights learnable prompt components using pooled features, and prompt interaction fuses the resulting prompt with features before transformer processing. The L1 loss compares restored and ground-truth images, and optimization uses Adam.Source: 2) Architecture · pp. 2–3

PHW C

The Lottery Ticket Hypothesis proposes that a sparse subnetwork can train from its original initialization to match or exceed the dense network with fewer parameters and iterations. Layer-wise pruning applies a fixed share of low-magnitude weights independently within each layer.

B. Lottery Ticket Hypothesis

  • A winning ticket is a sparse subnetwork trained in isolation from its original initialization that achieves at least the dense network’s test accuracy with no more iterations and far fewer parameters.The hypothesis represents the subnetwork with a binary mask applied to the dense network’s initial parameters.Source: B. Lottery Ticket Hypothesis · p. 3

1) Layer-wise Pruning:

  • Layer-wise pruning independently removes a fixed percentage of the smallest-magnitude weights in each layer.The approach distributes sparsity uniformly across layers.Source: 1) Layer-wise Pruning · p. 3
  • The MIR-L optimization algorithm takes a dense network, pruning rate, training and warmup epochs, batch size, training samples, learning-rate settings, and target sparsity as inputs.Its output is a trained sparse network.Source: 1) Layer-wise Pruning · p. 3
  • The algorithm initializes a binary mask and masked parameters, then trains for the specified epochs using linear warmup followed by cosine annealing.The supplied algorithm excerpt begins its iterative pruning loop after this training step.Source: 1) Layer-wise Pruning · p. 3

PHW C

MIR-L masks pruned weights during training, then removes low-magnitude weights and resets the survivors to their original initialization. Iterations continue until the target sparsity is reached.

1) Layer-wise Pruning:

  • After training, MIR-L sets a pruning threshold at the p-th percentile of the magnitudes of currently active weights.The gradients and updated weights are masked during training; parameters are updated via Adam.Source: PHW C · p. 3
  • The updated mask retains weights at or above the threshold, and the remaining active weights are reset to their initial values.The algorithm repeats pruning until the target sparsity is reached.Source: PHW C · p. 3
  • In layer-wise pruning, the output layer is pruned at half the rate because it has fewer parameters and aggressive pruning can yield diminishing returns earlier.This adjustment is described specifically for the output layer.Source: PHW C · p. 3

2) Global Pruning:

Global pruning selects low-magnitude weights across the network rather than separately by layer, avoiding bottlenecks in smaller layers. MIR-L iteratively trains and prunes the multi-task model, resetting surviving weights to their initial values until the target sparsity is reached.

2) Global Pruning:

  • Global pruning removes a fixed percentage of the smallest-magnitude weights across the network, avoiding bottlenecks from uniformly pruning smaller layers and identifying smaller winning tickets in networks with imbalanced layer sizes.The passage describes this approach as particularly effective in deeper networks.Source: 2) Global Pruning · p. 3

C. MIR-L Optimization Algorithm

  • MIR-L applies the Lottery Ticket Hypothesis to the multi-task restoration model to obtain a sparse model intended to be equally or more effective.The model is optimized with an L1 reconstruction loss and repeatedly trained and pruned until reaching the target sparsity.Source: C. MIR-L Optimization Algorithm · pp. 3–4
  • The pruning threshold is computed independently for each layer in layer-wise pruning, but across all layers in global pruning.Source: C. MIR-L Optimization Algorithm · p. 4

IV. EXPERIMENTAL EVALUATION

The evaluation combines denoising, deraining, and dehazing datasets to assess a unified model, using PSNR and SSIM as image-restoration metrics. PSNR is calculated from the maximum pixel value and the mean squared error between corresponding image pixels.

A. Datasets

  • The all-in-one evaluation combines denoising, deraining, and dehazing datasets for training and testing a unified model.Training uses 5% of BSD400 and WED, 10% of Rain100L pairs, and 3% of OTS pairs; testing uses Color BSD68, Urban100, Rain100L, and SOTS.Source: A. Datasets · p. 4

B. Evaluation Protocol

  • PSNR and SSIM are used to assess restoration quality, providing measures of reconstruction fidelity and perceptual similarity.Source: B. Evaluation Protocol · p. 4

1) Peak Signal-to-Noise Ratio (PSNR):

2) Structural Similarity Index Measure (SSIM):

SSIM measures perceived image quality through structural information, luminance, and contrast. The section presents its formula and explains why SSIM complements PSNR.

2) Structural Similarity Index Measure (SSIM):

1) Implementation Details:

Experiments use a common training protocol and compare proposed lottery-ticket pruning variants with restoration baselines across several tasks. All examined models are retrained using public implementations and the stated datasets.

1) Implementation Details:

  • Training uses 120 epochs, 90% target sparsity over 15 pruning steps, and 20% pruning per step, with Adam, L1 loss, and 64×64 patches.Runs use an NVIDIA A40 with PyTorch 2.5.1, batch size 8, learning rate 2 × 10^-4, and 15 warmup epochs; smaller datasets are duplicated and augmented.Source: 1) Implementation Details · p. 4

2) Examined Models:

  • The cited baselines include MSPFN, EPDN, AirNet, Restormer, and FFDNet, described respectively as networks for deraining, dehazing, all-in-one restoration, high-resolution restoration, and denoising.Source: 2) Examined Models: • · pp. 4–5
  • One-shot magnitude and random baselines prune 30% or 70% of weights; the 30% variants receive an additional 5% of training epochs for fine-tuning.Source: 2) Examined Models: • · p. 5
  • For a fair comparison, all examined models are retrained using publicly available implementations, the described datasets, and 64×64 input patches.Source: 2) Examined Models: • · p. 5

D. Experimental Results

Across single-task benchmarks, the proposed methods reduce parameters while preserving restoration performance, with global pruning consistently outperforming layer-wise pruning. The reported comparisons span deraining, dehazing, and denoising.

D. Experimental Results

  • Iterative pruning with weight resets outperforms one-shot pruning at high sparsity, while allowing surviving weights to relearn and recover performance.The method repeatedly removes low-magnitude weights and resets survivors to their original values; one-shot baselines show steep performance drops at high sparsity.Source: D. Experimental Results · p. 5
  • MIR-L-LW and MIR-L-G reduce parameters to 4.7M while preserving high performance across single-task deraining, dehazing, and denoising benchmarks.MIR-L-G consistently outperforms MIR-L-LW in all reported single-task settings.Source: D. Experimental Results · p. 5, Table II · p. 6
  • As trainable parameters decrease, global pruning generally preserves PSNR better than layer-wise pruning, but deraining has a high-sparsity exception.The figure plots PSNR against trainable parameter count; the text attributes the deraining drop to essential weights being pruned by the global magnitude criterion.Source: D. Experimental Results · p. 6, Figure 1 · p. 6

V. CONCLUSION

MIR-L applies iterative pruning and weight resets to discover sparse subnetworks for deraining, dehazing, and denoising. The authors report up to 90% fewer trainable parameters while maintaining high benchmark performance.

V. CONCLUSION

  • MIR-L targets deraining, dehazing, and denoising with iterative removal of low-magnitude weights and resetting surviving weights to their initial values.Source: §V · pp. 6–7
  • MIR-L reduces trainable parameters by up to 90% across single-task and multi-task settings while maintaining high benchmark performance.The conclusion frames lower model size and computational requirements as beneficial for web platforms and less powerful client devices.Source: §V · p. 7
  • The authors identify more sophisticated pruning criteria and extending the method to super-resolution as future directions.Source: §V · p. 7

Related papers

Loading 2510.14463v1…