Structural Similarity Filter Pruning of Frequency Domain Edge for Convolutional Neural Networks
- Resource Type
- Conference
- Authors
- Liu, Yajun; Fan, Kefeng; Wu, Dakui; Zhou, Wenju
- Source
- 2024 4th International Conference on Consumer Electronics and Computer Engineering (ICCECE) Consumer Electronics and Computer Engineering (ICCECE), 2024 4th International Conference on. :148-152 Jan, 2024
- Subject
- Communication, Networking and Broadcast Technologies
Components, Circuits, Devices and Systems
Robotics and Control Systems
Signal Processing and Analysis
Analytical models
Frequency-domain analysis
Image edge detection
Computational modeling
Feature extraction
Information filters
Frequency conversion
Discrete wavelet transform
edge information
structural similarity
filter pruning
- Language
Filter pruning has become one of the driving forces behind the deployment of convolutional neural networks (CNNs) in mobile devices due to its effective compressibility and software and hardware universality. Existing pruning methods are based on analyzing the inherent parameters of the network model in the spatial domain and searching for the redundant parameters present. This approach ignores the correlation of the model structure in the frequency domain. In this paper, filter pruning is performed with the help of discrete wavelet transform (DWT) and edge structure similarity of feature maps to realize the readability of the model structure in the frequency domain. The edge information is an important part of an image. In addition, the WT provides a multi-resolution representation of image edge and can analyze edge at a finer scale. In this paper, the frequency domain transformation of DWT is used to convert the spatial feature map to the frequency domain in order to accurately obtain the edge structural components of the feature map. The structural similarity (SSIM) is also used to calculate the similarity between feature map edges, which in turn prunes the filter based on the feature map importance score. Experiments on two image classification datasets demonstrate the feasibility of the method proposed in this paper.