A novel pooling method for CNN model based on discrete cosine transform

Aqeel Mohsin Hamad, Naofal Mohamad Hassein

Abstract


Deep learning can be used to learn huge volume of data, which will be processed through hidden layers and according to the number of hidden layers ,filter size and numbers and the required computation cost is increased because of the size of raw data, this problem can be avoided by using pooling techniques, different method s are proposed to extract the basic features of the signal instead of all signal, but unfortunately this operation may introduce some noise or omission because of elimination important data from the signal. In this paper, A novel pooling method are proposed based on discrete cosine transform , this method is utilized DCT technique to reduce spatial redundancy of image by transform the spatial domain into frequency domain , which can preserve the most significant image information from the other coefficients, which represents the other details information of the image, so discard these less important coefficients. Its effect will be slight and this can reduce the eliminated information as compared with other methods. After applying DCT, we crop the most significant coefficients to be used in the reconstructed data by applying inverse DCT . then the result is combined in different methods with Max pooling and average pooling methods, this new structure can reduce the effect of discarding most important information and reduce the drawbacks of average and Max. pooling method. لإhe results are proved that our proposed methods are outperformed some standard methods and can be used in more application.

Keywords


Deep learning, CNN, DCT, PDCTM, SDCTM

Full Text:

PDF


DOI: http://dx.doi.org/10.21533/pen.v9i3.2233

Refbacks

  • There are currently no refbacks.


Copyright (c) 2021 Authors

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.

ISSN: 2303-4521

Digital Object Identifier DOI: 10.21533/pen

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License