Scholarlink

The Fashion-MNIST data functions as a novel benchmark for various models, comprising 70,000 grayscale images of apparel—60,000 designated for training and 10,000 for testing. Each image has dimensions of 28 × 28 pixels, serving as a direct substitute for the traditional MNIST dataset, which comprises handwritten numbers. This study examines the efficacy of different deep learning models, such as Vision Transformers (VIT), Self-Attention CNNs, and Dense Net, in categorizing Fashion-MNIST photos into ten separate apparel categories. The dataset is organized with the initial column comprising class labels, while the subsequent 784 columns denote pixel values from 0 to 255, reflecting pixel intensity. We initiate the dataset preparation by normalizing and segregating pixel values from class labels, employing frameworks such as Kera’s and NumPy. Subsequently, we develop and assess convolutional neural networks (CNNs) for image classification, utilizing the distinct benefits of each model design. Vision Transformers, employing self-attention mechanisms, have exceptional proficiency in collecting spatial relationships in images, rendering them especially effective for intricate datasets. In contrast, Dense Net demonstrates efficiency via densely coupled layers that promote feature reuse and improve gradient flow. Model performance is meticulously evaluated using conventional measures such as Accuracy (CA), F1-score, Precision, Recall, and Area Under the Curve (AUC). Our findings indicate that although Vision Transformers have superior accuracy on intricate datasets, Dense Net exhibits competitive performance with a reduced number of parameters. This research provides significant insights into the utilization of advanced deep learning methodologies for picture classification, emphasizing the necessity of model selection according to dataset attributes and computing efficiency. The study seeks to improve comprehension of classification models and their appropriateness for practical applications in fashion and retail analytics.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Sartorial Sophistication: A Convolution Neural Network Odyssey in Fashion MNIST Image Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kaarthikeya Kammula,
  • Gunasekhar Devineni,
  • Sai Yaswanth Reddy Suram,
  • Shankar Chowdary Jampani

摘要

The Fashion-MNIST data functions as a novel benchmark for various models, comprising 70,000 grayscale images of apparel—60,000 designated for training and 10,000 for testing. Each image has dimensions of 28 × 28 pixels, serving as a direct substitute for the traditional MNIST dataset, which comprises handwritten numbers. This study examines the efficacy of different deep learning models, such as Vision Transformers (VIT), Self-Attention CNNs, and Dense Net, in categorizing Fashion-MNIST photos into ten separate apparel categories. The dataset is organized with the initial column comprising class labels, while the subsequent 784 columns denote pixel values from 0 to 255, reflecting pixel intensity. We initiate the dataset preparation by normalizing and segregating pixel values from class labels, employing frameworks such as Kera’s and NumPy. Subsequently, we develop and assess convolutional neural networks (CNNs) for image classification, utilizing the distinct benefits of each model design. Vision Transformers, employing self-attention mechanisms, have exceptional proficiency in collecting spatial relationships in images, rendering them especially effective for intricate datasets. In contrast, Dense Net demonstrates efficiency via densely coupled layers that promote feature reuse and improve gradient flow. Model performance is meticulously evaluated using conventional measures such as Accuracy (CA), F1-score, Precision, Recall, and Area Under the Curve (AUC). Our findings indicate that although Vision Transformers have superior accuracy on intricate datasets, Dense Net exhibits competitive performance with a reduced number of parameters. This research provides significant insights into the utilization of advanced deep learning methodologies for picture classification, emphasizing the necessity of model selection according to dataset attributes and computing efficiency. The study seeks to improve comprehension of classification models and their appropriateness for practical applications in fashion and retail analytics.