翻译成中文：We evaluate the performance of gUNet on four image dehazing datasets, and the results show that gUNet is comparable to or even better than stateof-the-art methods with a much smaller overhead. More importantly, we performed large-scale ablation studies and showed that the performance gains of image dehazing networks go mainly from the attention mechanism, nonlinear activation function, extraction of global information, normalization layer, and the number of training epochs.

我们评估了gUNet在四个图像去雾数据集上的性能，结果显示gUNet与最先进方法相当甚至更优，且开销更小。更重要的是，我们进行了大规模的消融研究，并表明图像去雾网络的性能提升主要来自于注意机制、非线性激活函数、全局信息提取、归一化层以及训练轮数。