LMWP-Net: Lightweight Multi-scale Weight Pruning Network for Salient Object Detection

This repository provides the official PyTorch implementation of the paper "Lightweight Multi-scale Weight Pruning Network (LMWP-Net) for Salient Object Detection".

📖 Introduction

Salient object detection (SOD) is fundamental to computer vision, yet deep learning approaches often suffer from high computational costs, limiting deployment on resource-constrained devices.

We propose a Lightweight Multi-scale Weight Pruning Network (LMWP-Net) to balance high performance with low complexity. LMWP-Net employs an encoder-decoder architecture featuring two key components:

1. Multi-scale Weight Pruning Module (MWPM): Extracts multi-scale features and effectively reduces redundant information via dynamic weight pruning.
2. Multi-scale Attention Fusion Module (MAFM): Enhances encoder features explicitly through spatial/channel attention mechanisms and fuses them with decoder features under the guidance of high-level semantics.

Extensive experiments on 5 public datasets demonstrate that LMWP-Net consistently outperforms state-of-the-art lightweight methods, offering a superior efficiency-to-effectiveness ratio for real-time applications.

✨ Highlights

• High Efficiency & Real-time Speed: Achieves an impressive inference speed (83.3 FPS on RTX 3090) with only 3.03M parameters and 0.5G FLOPs.
• Redundancy Reduction: The proposed MWPM dynamically discards task-irrelevant background weights.
• Robust Feature Integration: The MAFM aligns spatial and semantic details, mitigating background noise.

🏗️ Architecture

(Please place your network architecture image in the assets/ folder and link it here)

📊 Quantitative Results

Comparison with state-of-the-art lightweight SOD methods and backbones:

Method	Params (M)	FLOPs (G)	FPS	ECSSD ($F_\beta \uparrow$)	DUTS-TE ($F_\beta \uparrow$)	HKU-IS ($F_\beta \uparrow$)
HVPNet	1.23	1.1	26	0.925	0.815	0.915
SAMNet	1.33	0.5	44	0.925	0.812	0.915
MobileNetV2	2.37	0.8	446	0.905	0.798	0.890
LMWP-Net (Ours)	3.03	0.5	83.3	0.932	0.827	0.916

(Please place your qualitative comparison image in the assets/ folder and link it here)

⚙️ Get Started

1. Requirements

• Python $\ge$ 3.8
• PyTorch $\ge$ 1.9.0
• torchvision
• numpy, opencv-python, pillow

Install dependencies:

pip install -r requirements.txt