An improved MixDehazeNet for image dehazing via cross-scale interaction and structure-aware convolution

To enhance the imaging quality and target recognition capability of polarized images under low-altitude adverse weather conditions, this paper proposes an improved dehazing network for polarized images based on the original MixDehazeNet architecture. The proposed network incorporates a Lightweight Cross-Scale Module (LCSM) and an Efficient Structure- Aware Convolution (ESAC) module. Built upon the U-Net framework, the network leverages the inherent advantages of polarization imaging in scattering suppression and contrast enhancement, while optimizing its architecture for highresolution image processing. The ESAC module adopts a mixture-of-experts mechanism to adaptively model diverse structural features, whereas the LCSM module enables efficient interaction and fusion of multi-scale features. The proposed model contains as few as 4.4 million parameters, offering excellent deployment efficiency and application potential. Compared with existing dehazing networks such as Focal-Net and MixDehazeNet, the proposed approach achieves more effective haze removal and significantly alleviates issues such as detail loss.