Abstract
Image dehazing technology is widely utilized in outdoor environments, especially in precision agriculture, where it enhances image quality and monitoring accuracy. However, conventional dehazing methods have exhibited limited performance in complex outdoor conditions, necessitating the development of more advanced models to address these challenges. This paper proposes DUNet, a high-performance image dehazing model that is well-suited for outdoor smart agriculture applications. In this study, we first introduce a novel hybrid convolution block, MixConv, designed to fully extract detailed feature information from images. Secondly, by incorporating the atmospheric scattering model, we propose a dehazing feature extraction unit, DFEU, integrated between the encoder and decoder, to establish a mapping relationship between hazy and haze-free images in the feature space. Finally, the SK fusion mechanism dynamically fuses feature maps extracted from multiple paths. To evaluate the dehazing performance of DUNet, we constructed a dataset consisting of 1,978 pairs of hazy UAV images of paddy fields. DUNet achieved a PSNR of 36.0206 and an SSIM of 0.9946 on this dataset. We further validated DUNet's performance on a remote sensing dataset, achieving a PSNR of 37.2887 and an SSIM of 0.9933. Experimental results demonstrate that, compared to other well-established image dehazing models, DUNet offers superior performance, confirming its potential and feasibility for outdoor smart agriculture dehazing tasks.