Scholarlink

<p>Deep learning-based image classification remains challenging for low-resolution images due to feature loss and limitations in representation quality. This study introduces the guided convolutional variational autoencoder (GC-VAE) that integrates the variational autoencoder (VAE) framework with a classification network to enhance the learning of robust latent representations for improved low-resolution image classification. By leveraging a structured latent space, GC-VAE improves feature separability, enhancing classification performance while mitigating the loss of image details. A hybrid loss function is proposed by combining reconstruction loss, Kullback–Leibler divergence, and classification loss to optimize both generative and discriminative learning. The proposed GC-VAE model is evaluated on the benchmark CIFAR-10 dataset, achieving an accuracy of 82.95%, surpassing conventional CNN-based classifiers. Experimental results and visualizations demonstrate the effectiveness of the GC-VAE model in handling low-resolution image classification tasks. Furthermore, the proposed approach has significant practical relevance across multiple domains. In medical diagnostics, it facilitates disease detection from low-resolution medical images such as X-rays and retinal scans. In remote sensing, it supports land-use and object classification from coarse-resolution satellite imagery. In surveillance systems, it enhances object recognition and activity detection from low-quality video feeds. These application scenarios highlight the robustness and adaptability of the GC-VAE model in real-world low-resolution imaging environments.</p>

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

Guided Convolutional Variational Autoencoder for Learning Latent Representations in Low-Resolution Image Classification

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Jyostna Devi Bodapati,
  • Reshma Shaik,
  • Veeranjaneyulu N

摘要

Deep learning-based image classification remains challenging for low-resolution images due to feature loss and limitations in representation quality. This study introduces the guided convolutional variational autoencoder (GC-VAE) that integrates the variational autoencoder (VAE) framework with a classification network to enhance the learning of robust latent representations for improved low-resolution image classification. By leveraging a structured latent space, GC-VAE improves feature separability, enhancing classification performance while mitigating the loss of image details. A hybrid loss function is proposed by combining reconstruction loss, Kullback–Leibler divergence, and classification loss to optimize both generative and discriminative learning. The proposed GC-VAE model is evaluated on the benchmark CIFAR-10 dataset, achieving an accuracy of 82.95%, surpassing conventional CNN-based classifiers. Experimental results and visualizations demonstrate the effectiveness of the GC-VAE model in handling low-resolution image classification tasks. Furthermore, the proposed approach has significant practical relevance across multiple domains. In medical diagnostics, it facilitates disease detection from low-resolution medical images such as X-rays and retinal scans. In remote sensing, it supports land-use and object classification from coarse-resolution satellite imagery. In surveillance systems, it enhances object recognition and activity detection from low-quality video feeds. These application scenarios highlight the robustness and adaptability of the GC-VAE model in real-world low-resolution imaging environments.