高级检索

融合引导滤波的无监督图像超像素快速生成

Unsupervised Fast Image Superpixel Generation With Guided Filtering

  • 摘要: 深度学习已成为图像超像素分割任务的主流方法之一, 其中以有监督学习为主, 其性能往往依赖于大量的标记数据. 同时, 超像素作为一种常用的预处理步骤,其速度与准确性是算法性能的关键指标之二。因此,本文拟提出一种无监督、保持准确性的图像超像素快速生成模型. 首先, 利用引导滤波设计下采样-上采样的超像素快速生成框架以提高模型的推理速度; 然后, 采用多尺度注意力机制来提高超像素分割的准确性, 充分挖掘图像信息, 设计鲁棒的损失函数, 实现无监督的图像超像素生成模块; 最后采用顺序训练策略减少模型对大量训练数据的依赖. 在公开数据集BSDS500与DRIVE上测试的实验结果表明, 本文提出的模型拥有与有监督算法可比的准确性, 且优于已有无监督方法, 同时耗时最短.

     

    Abstract: Deep learning has become one of the mainstream methods for image superpixel segmentation, most of which are supervised. The performance often depends on a large number of labeled data. Moreover, as a commonly used preprocessing procedure, speed and accuracy are two key indicators of the performance of superpixel segmentation algorithms. Hence, this paper proposes an unsupervised model to generate superpixels fast with accuracy. Firstly, using guided filtering to design a downsampling-upsampling framework to improve the inference speed of suprpixel segmentation model. Then, multi-scale attention mechanism is adopted to improve the accuracy of superpixel segmentation. At the same time, a robust loss function is defined based on image information. In this way, an unsupervised image superpixel generation module was proposed; Finally, a sequential training strategy is adopted to reduce the model's dependence on a large amount of training data. The experimental results tested on the public datasets BSDS500 and DRIVE show that the proposed model has comparable accuracy with supervised algorithms, outperforms existing unsupervised methods, and achieves fast performance.

     

/

返回文章
返回