Similarity Based Filter Pruning for Efficient Super-Resolution Models
Chu Chu, Li Chen, Zhiyong Gao · 2020
Recently, deep convolutional neural networks have achieved significant success on image super-resolution tasks. Models with more complex structure typically contribute to better accruacy while at higher computational cost, which impeded the deployment of them on edge devices with limited resource. In this paper, we propose a similarity based filter pruning method to obtain efficient super-resolution models. Specifically, for every two convolutional filters, we determine their similarity by calculating the Euclidean distance between them and obtain a correlation matrix for the entire filter set. By pruning filters that are highly similar to all other ones, we can obtain a compact model without significant accuracy drop. In addition, after the weights initialization, we set the filters to be pruned to zero, and complete training process according to the obtained mask, which reduce the dependence on the pre-trained model and eliminate the fine-tuning step. Compared with existing pruning method, our experiments achieve better performance on super-resolution benchmarks.