Fuzzy Edge-Detection as a Preprocessing Layer in Deep Neural Networks for Guitar Classification

• Autor: Torres, Cesar ; Gonzalez, Claudia I. ; Martinez, Gabriela E.
• Assuntos: Accuracy ; Algorithms ; Artificial intelligence ; Brain research ; Classification ; Color imagery ; Computer architecture ; Computer vision ; convolutional neural networks ; Diabetes ; Diabetic retinopathy ; fuzzy edge-detection ; Fuzzy sets ; guitar recognition ; Guitars ; Image classification ; Inventory ; Magnetic resonance imaging ; Morphology ; Neural networks ; Preprocessing ; Sensors ; Vision systems
• É parte de: Sensors (Basel, Switzerland), 2022-08, Vol.22 (15), p.5892
• Notas: ObjectType-Article-1
  SourceType-Scholarly Journals-1
  ObjectType-Feature-2
  content type line 23
  
• Descrição: Deep neural networks have demonstrated the capability of solving classification problems using hierarchical models, and fuzzy image preprocessing has proven to be efficient in handling uncertainty found in images. This paper presents the combination of fuzzy image edge-detection and the usage of a convolutional neural network for a computer vision system to classify guitar types according to their body model. The focus of this investigation is to compare the effects of performing image-preprocessing techniques on raw data (non-normalized images) with different fuzzy edge-detection methods, specifically fuzzy Sobel, fuzzy Prewitt, and fuzzy morphological gradient, before feeding the images into a convolutional neural network to perform a classification task. We propose and compare two convolutional neural network architectures to solve the task. Fuzzy edge-detection techniques are compared against their classical counterparts (Sobel, Prewitt, and morphological gradient edge-detection) and with grayscale and color images in the RGB color space. The fuzzy preprocessing methodologies highlight the most essential features of each image, achieving favorable results when compared to the classical preprocessing methodologies and against a pre-trained model with both proposed models, as well as achieving a reduction in training times of more than 20% compared to RGB images.
  
• Editor: Basel: MDPI AG
  
• Idioma: Inglês