<p>At present, the image dehazing method based on deep learning is difficult to take into account the global features and local details when dealing with uneven haze, resulting in the loss of details and the decrease of contrast in the restored image. Therefore, a multi-scale contrast-aware dynamic feature fusion dehazing method based on transformer is proposed. The transformer in the network encoder consists of a multi-head proxy attention block (MPAB), a residual multi-scale attention block (RMAB), and a feed-forward network (FFN). Among them, MPAB introduces a multi-head proxy attention mechanism with low computational overhead to extract multi-scale global features, so as to significantly improve the inference efficiency while ensuring the dehazing quality (FLOPs is 35.6G, running time is 0.095&#xa0;s), showing excellent computational performance. RMAB uses the residual structure to capture local features and context information to enhance the ability of multi-scale feature modeling. Between the encoder and the decoder, a channel-spatial dynamic fusion block (CSDF) is introduced to integrate low-level and high-level features through a dynamic fusion mechanism to prevent information loss. At the same time, the adaptive contrast-aware enhancement block (ACEB) is added to the middle layer, and four modules with different expansion rates are stacked to adapt to the uneven distribution of haze in the real scene. The experimental results show that the method performs well on multiple synthetic and real fog map datasets and has both accuracy and efficiency. It can effectively solve the problem of detail feature loss and contrast reduction.</p>