Accelerating convolutional neural network models based on wavelet transform
Zihan Shen, Jianchuang Qu, Kaige Wang, Can Wu, Qing Li · 2024
Convolutional neural networks require a large amount of computing resources and time to achieve progress and development in computer vision tasks. Wavelet transform can provide multi-resolution features of images. By using wavelet transform to preprocess the images in the training set, the main information of the images can be preserved. The processed images can then be used as input for the neural network, significantly reducing the training time. By comparing different wavelet bases and orders, it was found that Bior wavelets showed the best acceleration effect, and the training time was significantly reduced. If the complexity of the model is appropriately increased, the training accuracy can be improved while the training time is reduced by 44% compared with the original time.