Adams-based hierarchical features fusion network for image dehazing

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional Neural Networks (CNNs), particularly Residual Networks (ResNets), are effective for image dehazing.
• ResNets' success is linked to their formulation similar to the Euler method for Ordinary Differential Equations (ODEs).
• Image dehazing can be viewed as an optimal control problem in dynamical systems.

Purpose of the Study:

• To propose a novel image dehazing method inspired by multi-step optimal control solvers.
• To leverage the stability and efficiency advantages of multi-step methods over single-step solvers like Euler.
• To introduce the Adams-based Hierarchical Feature Fusion Network (AHFFN) for enhanced image restoration.

Main Methods:

• Extended the multi-step Adams-Bashforth method into an Adams block for higher accuracy.
• Stacked Adams blocks to simulate the discrete approximation of optimal control in dynamical systems.
• Integrated Hierarchical Feature Fusion (HFF) and Lightweight Spatial Attention (LSA) into Adams blocks to form a new Adams module.

Main Results:

• The proposed Adams module effectively fuses hierarchical features and highlights spatial information.
• Experimental results on synthetic and real images show improved accuracy and visual quality.
• The AHFFN outperforms existing state-of-the-art image dehazing methods.

Conclusions:

• The AHFFN provides a new perspective on image restoration using optimal control principles.
• Multi-step optimal control methods offer advantages in stability and efficiency for image dehazing.
• The proposed method demonstrates significant improvements in both quantitative and qualitative evaluations.