Innovative solutions for convolutional neural network performance: A TRIZ-based reverse engineering approach

Abstract

Convolutional neural networks (CNNs) are widely used in computer vision for tasks like image classification and detection. These models work well when the number of image classes is small, but as the number of classes increases, accuracy tends to drop due to overfitting. There are several methods to address this issue, such as data augmentation, preprocessing, class weighting, transfer learning, and adjusting technical parameters. This study introduces a novel approach utilizing the theory of inventive problem-solving (TRIZ) methodology to systematically analyze and enhance these existing methods. Using reverse engineering, we deconstructed current solutions and aligned them with TRIZ principles to propose more innovative and effective approaches for improving CNN performance. The results show that TRIZ provides a structured and creative framework for solving accuracy decline issues in CNN models, offering the potential for broader applications in other machine learning architectures. © 2025, Society of Sytematic Innovation. All rights reserved.