Handwritten digits recognition using transfer learning. (March 2023)

• Record Type: Journal Article
• Authors: Azawi, Nidhal
• Is Part Of: Computers & electrical engineering. Volume 106(2023)
• Publication Date: 2023-03
• Subjects: Deep CNN models -- Recognition -- Verification -- Statistical information -- Intra-class refinement; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2023.108604
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: A new proposed algorithm for multi-class recognition with a one-step verification system. Detailed experimental analysis that identifies the effectiveness of the 13 developed deep CNNs models, as well as the hybrid ensemble model. A statistical approach for selecting only the most qualified models in recognizing patterns. A new method for approaching the deep learning intra-class refinement problem in multiclass recognition systems which substantially improves the accuracy for each class. Abstract: Researchers often focus on building models that maximize overall predictive accuracy. In practice, however, it can be important for a model to yield good accuracy with each class value. Toward this end, a new recognition with a one-step verification methodology is proposed. It emphasizes the accuracy of each class value. The proposed discriminative system constructs an ensemble using several deep Convolutional Neural Networks (CNNs) with the help of statistical information. To the best of our knowledge, this is the first ensemble model that combines many deep CNNs with a focus on maximizing the accuracy for each class, rather than just overall accuracy. Experimental results show that the demonstration models achieved accuracy in the range of 97.82% to 99.72% within only a few epochs, rivaling the state-of-the-art. These results indicate that the performance of the proposed approach substantially improves the intra-class correlation, leading to improved classification accuracy for each class. Graphical abstract: Image, graphical abstract