<p>In the domain of image defogging utilizing deep convolutional generative adversarial networks, the challenge of gradient vanishing is commonly encountered during training. This study proposes a novel image defogging algorithm based on generative adversarial networks (GAN) that employs a multi-terminal generator architecture. In deep convolutional neural networks, excessively deep layers can exacerbate the impact of gradient vanishing on backpropagation. To address this issue, a network connection strategy using cross-layer residual connections has been suggested; however, residual connections not only increase the complexity of the network but also make the training process more difficult. The generator in our proposed GAN is segmented into a multi-terminal processing mode and interconnected in a sequential manner. When training the discriminator, the loss is computed as the aggregate of the losses from all generators, while each generator is trained with the loss derived from individually misleading the discriminator. The outputs of the multiple generators are then fused at various scales to enhance defogging performance. The experimental results indicate that the multi-terminal generator GAN has a faster training speed compared to the residual connection GAN, and it achieves better defogging performance compared to other advanced defogging algorithms.</p>