LEFTNet: low-frequency enhanced feature transmission network for dual-branch collaborative single image dehazing
TL;DR
Comprehensive experiments on four benchmarks demonstrate that LEFTNet achieves competitive performance across the evaluated Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Average Mean Brightness Error (AMBE), Gradient Magnitude Similarity Deviation (GMSD), and Visual Saliency-Induced Index (VSI) metrics compared with recent dehazing methods.
Abstract
Image dehazing is a basic low-level vision task designed to recover high-fidelity images from hazed images. In autonomous driving, aerial surveillance, and outdoor scene recognition, image quality is significantly degraded by light attenuation due to scattering in the propagation medium, especially under hazy, foggy, or smoggy conditions. While remarkable advances have been made with existing deep learning methods, they remain constrained by over-smoothing, ineffective use of haze frequency, and insufficient consideration of multi-scale feature interactions. To meet these challenges, we introduce a novel single-image dehazing framework, called Low-Frequency Enhanced Feature Transmission Network (LEFTNet), which consists of four special modules: the Haze Suppression Block (HSB), the Dual-Branch Block (DBB), the Cross-Scale Fusion Block (CSFB), and the Channel Mining Mechanism (CMM). In the first stage, a hazy input image is decomposed into two complementary low-frequency representations: a haze-related component ( I haze L ) and a frequency-domain component ( I freq L ), which are jointly processed by the HSB for early, channel-selective haze attenuation. Then, the encoder extracts multi-scale hierarchical features, which are cleaned up at each stage by the Channel Mining module. The DBB's residual refinement enables intra-image global-to-local feature interactions, while the CSFB's learnable fusion coefficients enable inter-scale feature fusion. A dual fully connected bottleneck with Global Average Pooling (GAP) suppresses haze-dominant channels while enhancing structure-sensitive channels. Two symmetric branches ensure complementary representation learning, and a hybrid loss function comprising L1 loss and perceptual LSSIM loss assures both pixel-wise accuracy and perceptual fidelity. Comprehensive experiments on four benchmarks, RESIDE-6K, I-Haze, O-Haze, and NH-Haze, demonstrate that LEFTNet achieves competitive performance across the evaluated Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Average Mean Brightness Error (AMBE), Gradient Magnitude Similarity Deviation (GMSD), and Visual Saliency-Induced Index (VSI) metrics compared with recent dehazing methods.