Abstract: The speed and accuracy of superpixel generation are the key indexes to evaluate the performance of supervised image superpixel segmentation methods, which usually rely on a large number of labeled data. Therefore, an unsupervised and accurate fast image superpixel generation method is proposed. Firstly, guided filter is used to design a fast superpixel generation framework of down-sampling and up-sampling to improve the inference speed. Then, atrous spatial pyramid pooling and multi-scale attention mechanism are used to fully mine image information, and a robust loss function is designed to improve the accuracy of superpixel generation, achieving an unsupervised image superpixel generation method. Finally, the sequential training strategy is adopted to reduce the dependence of the model on a large number of training data. Experimental results on publicly available datasets BSDS500 and DRIVE show that the proposed method enhances boundary recall rates and achievable segmentation accuracy by approximately 1% and 2%, respectively, compared to existing unsupervised methods, while achieving a 50% speed improvement. Moreover, it exhibits comparable performance to supervised methods.