Paper
8 April 2024 Model compression algorithm and model recovery strategy based on channel importance
Zhixiong Zhang, Yun Wu
Author Affiliations +
Proceedings Volume 13090, International Conference on Computer Application and Information Security (ICCAIS 2023); 130902G (2024) https://doi.org/10.1117/12.3026123
Event: International Conference on Computer Application and Information Security (ICCAIS 2023), 2023, Wuhan, China
Abstract
In the field of deep learning, object detection algorithms, especially the RetinaNet model based on convolutional neural networks, have garnered considerable attention due to their excellent multi-scale object detection performance. However, improvements in network performance often come at the cost of sacrificing inference speed, particularly when making enhancements to the network structure. This paper focuses on mitigating this negative impact by optimizing the improved RetinaNet model using channel pruning techniques. Initially, we conducted a quantitative analysis to assess the importance of each channel within the network to determine a pruning strategy. Subsequently, the network was finely tuned based on the set pruning rate, effectively reducing the model’s parameter count and computational complexity. It is noteworthy that the pruning process may lead to a decrease in detection performance. To address this, we introduced a novel model recovery method based on channel importance to compensate for the loss in performance. Experimental results indicate that this method significantly improves the model’s inference speed while maintaining a relatively high detection accuracy. The achievements of this research not only provide a new enhancement path for the practicality of the RetinaNet model but also bring a fresh perspective to the field of deep learning model optimization.
(2024) Published by SPIE. Downloading of the abstract is permitted for personal use only.
Zhixiong Zhang and Yun Wu "Model compression algorithm and model recovery strategy based on channel importance", Proc. SPIE 13090, International Conference on Computer Application and Information Security (ICCAIS 2023), 130902G (8 April 2024); https://doi.org/10.1117/12.3026123
Advertisement
Advertisement
RIGHTS & PERMISSIONS
Get copyright permission  Get copyright permission on Copyright Marketplace
KEYWORDS
Data modeling

Matrices

Object detection

Education and training

Performance modeling

Machine learning

Mathematical modeling

Back to Top