Regularization approach is a key technique to improve the generalization ability of models. The current dropblock method removes information pixels on images by covering black pixels or patches of random noise. Such deletion is undesirable because it leads to information loss and inefficiency in the training period. Therefore, we propose to use better noise than vacancies for feature regularization, we use high-level feature map of the network as perturbations, and these high-level feature additions avoid losing useful information compared to vacancy additions, and this approach is easier to implement. It is experimentally demonstrated that our method has more boosting and generalization capability than dropblock.
Selfreg: Regularization Method via High-level Feature
2021-10-20
1230362 byte
Conference paper
Electronic Resource
English
Enhancing feature tracking with gyro regularization
British Library Online Contents | 2016
|Feature-enhancement of SAR images by Bayesian regularization
Tema Archive | 2005
|Part level transfer regularization for enhancing exemplar SVMs
British Library Online Contents | 2015
|Semi-supervised sparse feature selection based on multi-view Laplacian regularization
British Library Online Contents | 2015
|On the Length and Area Regularization for Multiphase Level Set Segmentation
British Library Online Contents | 2010
|