LWDA-Net: A Lightweight Dual-Attention Network for Single Image Dehazing
Zhuo Hua, Zhen Hua, Jinjiang Li · 2024
Single image dehazing remains a challenging problem in vision tasks. In this paper, we propose an end-to-end Lightweight Dual Attention Network (LWDA-Net) for direct recovery of haze-free images.The key part of the LWDA-Net architecture is the novel Dual Attention (DA) module, which takes into account completely different weighting information between samples and provides stronger regularisation.The DA module unequally handles feature maps with different sized kernels, offering more possibilities to handle different types of information, the DA module can dynamically modify the local receptive field size of neurons, making the model more flexible to adapt to the diversity of inputs. Attention-based LWDA-Net is a neural network architecture that dynamically learns feature weights from the DA module, prioritizing important features. Notably, it retains and conveys shallow information to deeper layers. Experimental results show that our proposed LWDA-Net outperforms previous single-image dehazing methods both quantitatively and qualitatively, and even outperforms them in the dehazing task on the SOTS dataset.