본문 바로가기
  • Home
Basic Search

Design of Convolution Neural Network Based On Solar Energy and Enhancement of Number Recognition

  • Journal of Knowledge Information Technology and Systems
  • Abbr : JKITS
  • 2019, 14(1), pp.93-101
  • DOI : 10.34163/jkits.2019.14.1.010
  • Publisher : Korea Knowledge Information Technology Society
  • Research Area : Interdisciplinary Studies > Interdisciplinary Research
  • Received : January 7, 2019
  • Accepted : February 8, 2019
  • Published : February 28, 2019

Kwon Oh-Sung 1

1공주교육대학교

Accredited

ABSTRACT

Applications of solar power are also increasingly common. In this paper, we explained the characteristics of our new discriminator based on solar energy and CNN for number recognition. Our CNN(Convolution Neural Network) is one in deep learning models with the function of human visual perception. The neural network model carries out the image feature matching process to recognize entered digital images. Our proposed CNN structure was designed with multiple hidden layers to retrieve the features of handwritten numeral images. The CNN structure performs 2D image convolution operation to extract noise, edge and stroke information from the numeral images. These feature information from the convolution layers is served as a new input stream of FCN(Fully Connected Network) for the classification of the features. The FCN performs a classification procedure based on the protypes already deployed for numeral image matching. We used MNIST(Modified National Institute of Standards and Technology) data set to achieve objective recognition measurement. The 60,000 samples of MNIST data set were applied for the learning process of our CNN recognizer, while we used the 10,000 samples for the performance measurement of our CNN recognizer. The measurement experiments were carried out after the repeating the training process of CNN only two times. The several misrecognition results were observed in our performance test, therefore we performed the additional training step to enhance the recognition rate of our CNN. Our additional method is the distortion of our input samples. we performed serveral distortion function such as a scaling, a rotation, and an elastic. In the final measurement of CNN recognizer, we observed that the recognition rate after the applying distortion process was improved by about 98.04%.

Statistics

  • 459 Viewed
  • 145 Downloaded

Tools

Issue List

Citation status

* References for papers published after 2023 are currently being built.

Basic Search Advanced Search