To address the issues of high computational complexity and large parameter size in convolutional neural network (CNN)-based image dehazing, this study proposes a lightweight dehazing network (LDNet). First, the atmospheric scattering model is reformulated to directly suppress haze noise, thereby reducing cumulative errors in intermediate variable estimation. Second, a reverse residual network module with an attention mechanism (RNAM) is designed to extract multi-scale features while emphasizing critical semantic information, effectively reducing model complexity and parameter size. Finally, a joint loss function combining L1 smoothing loss and multi-scale structure similarity (MS-SSIM) loss is used to improve reconstruction quality. The experimental results show that the proposed method outperforms existing approaches in terms of structural similarity and peak signal-to-noise ratio (PSNR) on synthetic datasets, while also achieving effective dehazing performance on realworld images. In addition, the model exhibits reduced parameter size and improved computational efficiency.