Related Experiment Videos
DMSH-Net: Depth-aware multi-scale hybrid vision network for image dehazing
Chenping Zhao1, Jun Li1, Yingjun Wang1
1School of Computer Science and Technology, Henan Institute of Science and Technology, Xinxiang, China.
Plos One
|August 4, 2026
Summary
A new network, DMSH-Net, effectively removes haze from single images by considering depth and scale. This Depth-Aware Multi-Scale Hybrid Vision Network improves image quality in challenging conditions.
Area of Science:
- Computer Vision
- Image Processing
- Artificial Intelligence
Background:
- Single-image dehazing is difficult due to depth-dependent and non-uniform haze.
- Existing methods struggle with spatially varying atmospheric conditions.
Purpose of the Study:
- To develop a robust single-image dehazing method.
- To improve haze removal accuracy by considering scene depth and multi-scale features.
Main Methods:
- Proposed DMSH-Net (Depth-Aware Multi-Scale Hybrid Vision Network).
- Introduced Convolutional Squeeze-and-Excitation Attention (CSEA) for joint channel and spatial modeling.
- Developed Nonlinear CSEA-coupled Residual Block (NCCRB) for adaptability to varying haze densities.
- Incorporated multi-scale dilated convolution for contextual aggregation.
Main Results:
- DMSH-Net demonstrated superior quantitative performance on standard benchmarks.
- Achieved excellent results in both full-reference and no-reference evaluations.
- Validated robustness in complex real-world dehazing scenarios.
Conclusions:
- DMSH-Net effectively addresses challenges in single-image dehazing.
- The network's design enables robust haze removal across diverse conditions.
- Implicitly captures haze variations through hierarchical recalibration and multi-scale aggregation.