Abstract: To address the limitations of traditional dehazing methods in complex haze scenarios—such as insufficient global modeling capacity, loss of local details, and poor adaptability to non-uniform haze—this paper proposes an adaptive image dehazing model based on Transformer architecture. By integrating multi-scale feature fusion and a dynamic gated attention mechanism, the model effectively balances global context modeling with local detail restoration. Specifically, a multi-scale Transformer encoder is designed, incorporating dilated convolutions to expand the receptive field and enhance high-frequency feature extraction. Furthermore, a Dynamic Gated Attention Module (DGAM) is introduced, which employs a dual-branch channel-spatial calibration strategy to adaptively fuse multi-scale features, improving robustness to non-uniform haze distribution. A hybrid loss function is also constructed, combining physical model constraints with perceptual loss to optimize reconstruction quality. Experiments conducted on multiple public datasets as well as a custom-built dataset demonstrate that the proposed method outperforms existing mainstream approaches in terms of image quality, particularly in recovering details under non-uniform haze conditions. Visualization results further confirm the model's advantages in color fidelity and edge sharpness, offering strong technical support for intelligent perception systems in low-visibility environments.