Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Pruning Overparameterized Multi-Task Networks for Degraded Web Image Restoration

Thomas Katraouras, Dimitrios Rafailidis

TL;DR

The paper tackles web image degradation by proposing MIR-L, an iterative Lottery Ticket–based pruning method for compressing overparameterized multi-task image restoration models. By comparing layer-wise and global pruning, MIR-L identifies sparse winning tickets that maintain or improve restoration performance with significantly fewer trainable parameters (down to 4.7M from 35.6M). Through an MIR-L optimization algorithm that trains and prunes in rounds while resetting active weights to initialization, the approach achieves up to 90% parameter reduction while preserving PSNR and SSIM across deraining, dehazing, and denoising benchmarks. The work demonstrates practical impact for web platforms, enabling faster, more accessible client-side restoration without sacrificing quality, and opens avenues for further efficiency gains via alternative pruning criteria and real-time tasks.

Abstract

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.

Pruning Overparameterized Multi-Task Networks for Degraded Web Image Restoration

TL;DR

The paper tackles web image degradation by proposing MIR-L, an iterative Lottery Ticket–based pruning method for compressing overparameterized multi-task image restoration models. By comparing layer-wise and global pruning, MIR-L identifies sparse winning tickets that maintain or improve restoration performance with significantly fewer trainable parameters (down to 4.7M from 35.6M). Through an MIR-L optimization algorithm that trains and prunes in rounds while resetting active weights to initialization, the approach achieves up to 90% parameter reduction while preserving PSNR and SSIM across deraining, dehazing, and denoising benchmarks. The work demonstrates practical impact for web platforms, enabling faster, more accessible client-side restoration without sacrificing quality, and opens avenues for further efficiency gains via alternative pruning criteria and real-time tasks.

Abstract

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.
Paper Structure (23 sections, 1 theorem, 2 equations, 1 figure, 2 tables, 1 algorithm)

This paper contains 23 sections, 1 theorem, 2 equations, 1 figure, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Magnitude Pruning.
  4. One-shot Pruning.
  5. Iterative Pruning
  6. Proposed Method
  7. Multi-Task Image Restoration Model
  8. Tasks
  9. Architecture
  10. Lottery Ticket Hypothesis
  11. Layer-wise Pruning
  12. Global Pruning
  13. MIR-L Optimization Algorithm
  14. Experimental Evaluation
  15. Datasets
  16. ...and 8 more sections

Key Result

Proposition 1

Consider a dense feed-forward neural network $f(x; \theta)$ with initial parameters $\theta_0 \sim \mathcal{D}_\theta$. Let $m \in \{0, 1\}^{|\theta|}$ be a binary mask that identifies the active connections in the subnetwork. The Lottery Ticket Hypothesis predicts that a mask $m$ does exist such th

Figures (1)

  • Figure 1: PSNR vs. trainable parameter count across progressive pruning steps. The x-axis denotes the number of trainable parameters, where a larger pruning step corresponds to fewer trainable parameters. MIR-L-G consistently maintains high performance as step (sparsity) increases, while MIR-L-LW experiences a sharp drop at higher pruning levels.

Theorems & Definitions (2)

  • Definition 1: Winning Ticket
  • Proposition 1