Superpixel is a local compact homogeneous region formed by adjacent pixels with similar characteristics. Compared with pixel features, replacing original pixels with superpixels can reduce the complexity of image processing algorithm. However, there are few researches on extracting superpixel features by deep neural networks. Aiming at the problem that the existing superpixel segmentation algorithm is not differentiable, which makes it unable to be integrated into the deep learning network. Based on SLIC, this paper converts the non-differentiable pixel-superpixel correlation calculation into differentiable operation, and proposes a differentiable superpixel segmentation method, which can extract superpixel features by using full convolution network. The new proposed superpixel segmentation algorithm has been tested on the BSDS 500 public dataset. The simulation results show it not only has excellent performance and higher accuracy, but also can used to learn other task-specific superpixels.
A novel improved SLIC superpixel segmentation algorithm
2022-10-12
1471285 byte
Conference paper
Electronic Resource
English
The SLIC-7 laser gyro inertial guidance system
Tema Archive | 1977
|Subsampling-based acceleration of simple linear iterative clustering for superpixel segmentation
British Library Online Contents | 2016
|Superpixel-based class-semantic texton occurrences for natural roadside vegetation segmentation
British Library Online Contents | 2017
|